1

When Geoscience Meets Generative AI and Large

Language Models: Foundations, Trends, and Future
Challenges
Abdenour Hadid, Tanujit Chakraborty, and Daniel Busby
arXiv:2402.03349v1 [physics.geo-ph] 25 Jan 2024

Abstract—Generative Artificial Intelligence (GAI) represents

an emerging field that promises the creation of synthetic data and
outputs in different modalities. GAI has recently shown impres-
sive results across a large spectrum of applications ranging from
biology, medicine, education, legislation, computer science, and
finance. As one strives for enhanced safety, efficiency, and sustain-
ability, generative AI indeed emerges as a key differentiator and
promises a paradigm shift in the field. This paper explores the
potential applications of generative AI and large language models
in geoscience. The recent developments in the field of machine
learning and deep learning have enabled the generative model’s
utility for tackling diverse prediction problems, simulation, and
multi-criteria decision-making challenges related to geoscience
and Earth system dynamics. This survey discusses several GAI
models that have been used in geoscience comprising generative
adversarial networks (GANs), physics-informed neural networks
(PINNs), and generative pre-trained transformer (GPT)-based
structures. These tools have helped the geoscience community in
several applications, including (but not limited to) data gener-
ation/augmentation, super-resolution, panchromatic sharpening,
haze removal, restoration, and land surface changing. Some
challenges still remain such as ensuring physical interpretation,
nefarious use cases, and trustworthiness. Beyond that, GAI
models show promises to the geoscience community, especially
with the support to climate change, urban science, atmospheric
science, marine science, and planetary science through their
extraordinary ability to data-driven modeling and uncertainty Fig. 1. The relation between large language models (LLMs), generative AI,
quantification. and other related learning schemes.

I. I NTRODUCTION the first idea of generative models dates back to the 1950s
Among the most fascinating topics that have recently at- with the introduction of Gaussian mixture models (GMMs) [1]
tracted the attention of researchers, enthusiasts, and even the and hidden Markov models (HMMs) [2]. One of the earliest
general public is the Generative Artificial Intelligence (GAI). applications of these tools was in the field of speech recogni-
The foremost ability of GAI lies in creating new content such tion [3]. After the rise of deep learning, the productivity and
as text, images, and audio, and hence fostering creativity. The capability of generative models have enhanced significantly.
main purpose of such models is to generate new samples Deep learning models are a class of AI models that can be
from what was already there in the training data. This new either discriminative or generative. Discriminative models are
technological development has pushed the boundaries of what typically trained on a large dataset of labeled samples, aiming
machines can achieve. As illustrated in Fig. 1, generative AI is to discover the relationship between the features of the samples
a subset of deep learning methods that can work with both la- and the given labels. Once a discriminative model is trained,
beled and unlabeled data of varied modalities using supervised, it can be utilized to predict the label of any new sample.
unsupervised, or semi-supervised learning schemes. However, A generative model, on the other hand, creates new sample
instances based on a learned probability distribution of existing
A. Hadid is with the Sorbonne Centre for Artificial Intelligence, Sorbonne samples.
University Abu Dhabi, UAE. (e-mail: [email protected]). To create new content using prompt engineering, GAI
Corresponding Author: T. Chakraborty is with the Department of Sci-
ence and Engineering, Sorbonne University Abu Dhabi, UAE, and also models are built using large-scale datasets during training. This
with the Sorbonne Centre for Artificial Intelligence, Paris, France. (e-mail: boosts GAI models to produce excellent results by recognizing
[email protected]). patterns in the data based on a probability distribution and then
D. Busby is with TotalEnergies, PAU, France. (e-mail:
[email protected]). creating “similar” patterns when prompted. GAI can run on
A. Hadid and T. Chakraborty have equal contribution. different models that use different mechanisms to do training
 2

and output new content. These include, for example, generative machine learning algorithms, and limiting generalization.
adversarial networks (GANs) [4], diffusion models [5], [6], Despite the rapid development, GAI is still in its early
transformers [7], variational autoencoders (VAEs) [8] and stages, as the success of these models is clearly tied to the
Large Language Models (LLMs) [9], [10]. One of the major quantity of the training data associated with the foundation
breakthroughs in GAI is the introduction of GANs in 2014 [4] model. Also, from the trustworthiness and fairness viewpoints,
in which one network generates content (generative model) some serious concerns have been raised about the potential
and the other helps in figuring out whether it is an authentic emergence of “super-intelligent machines” without adequate
sample or not (discriminative model). The next breakthrough safeguards. Generative AI can indeed unintentionally produce
came with the idea of Transformers [7] in the field of natural biased or incorrect information due to biases in its training
language processing (NLP) and computer vision (CV), fol- data. Moreover, generative models usually require a large
lowed by the creation of BERT [11] and GPT [12]. Most amount of high-quality, unbiased data to operate. Other equally
recently, the ChatGPT model was introduced, and the first important issues include the latency for generating high-
version was created by OpenAI [13]. These GAI models quality samples, the massive computing power that is needed
have shown impressive capabilities across a broad spectrum of to train generative models, and the maxim of “garbage in
tasks, much beyond showcasing its generative power in writing and garbage out”. All these issues are important to be taken
emails and reports. It has, in fact, attracted increasing attention into consideration to effectively capitalize and guarantee the
and has been recently deployed in numerous downstream beneficial use of generative AI in the field of geoscience
applications, encompassing diverse fields such as biology and and beyond. This survey discusses the potential of generative
medicine [14]–[20], education [21]–[24], legislation [25]–[28], AI in geoscience, providing some guidance, highlighting the
computer programming [29]–[34], and finance [35]–[40]. New challenges, and pointing out some future research directions.
applications of generative AI are appearing every day. The rest of this article is structured as follows. Section II dis-
The rapid progress in artificial intelligence has been a cusses and enumerates the potential of generative AI and large
game-changer across various industries, and geoscience is no language models in geoscience. Then, an overview of existing
exception. The integration of GAI in geoscience is revolu- GAI models and tools in geoscience are thoroughly presented
tionizing the way geoscientists interpret and understand the in Section III. Section IV describes some useful benchmark
Earth’s complex, interactive, and multiscale processes [41]. datasets and resources for geoscience GAI. Discussion and
This fusion of technology and science is proving to be a conclusions are drawn in sections VI and VII, respectively.
catalyst for innovation and progress, offering unprecedented
opportunities for breakthroughs in the field of geoscience.
II. G EOSCIENCE IN THE E RA OF G ENERATIVE AI
The availability of massive volumes of geo and Earth system
data is encouraging the applications of machine learning and This section discusses some potential applications of
deep learning tools for data-driven geoscience, Earth science, generative AI in geoscience. The most natural use is perhaps
and remote sensing [42]. In such scenarios, AI is proven through generating “new” samples for training the AI-based
to handle the complexity and ambiguity of geophysical data models. This is useful in applications like reservoir modeling
better than classical geostatistical methods. Geophysical data is in which geoscientists are often required to deal with several
often noisy and incomplete, making it challenging to interpret. samples of the subsurface and develop probabilistic models to
AI algorithms, particularly deep learning models, are designed help with the subsequent decision-making process. Examples
to handle such complexities. They can identify patterns and of current attempts at using generative AI in geoscience
relationships in the data that humans might misread, leading include but are not limited to, the reconstruction of 3D
to more accurate and reliable interpretations. For instance, AI porous media [45], the generation of geologically realistic
can predict seismic activities by analyzing patterns in historical 3D reservoir facies using deep learning of sedimentary
data, thereby helping in disaster management [43]. Further- architecture [46], and seismic data augmentation [47] using
more, AI can assist in resource exploration by identifying seismic waveform synthesis and GANs. More details on
optimal sites for oil and gas extraction [44]. This not only seismic data augmentation modeling (”SeismoGen”) are given
increases efficiency but also reduces the environmental impact in Section III-A as a use case of using generative AI in
of these activities. Perhaps one of the most recent and notable geoscience. Another use case of large language models (called
examples of generative AI is ChatGPT - a language model K2) for understanding and answering geoscience questions is
developed by OpenAI (https://openai.com/blog/chatgpt). It is discussed in Section III-B. The K2 model is claimed to be
trained on a diverse range of texts and is capable of generating able to generate new geoscience ideas as well. As illustrated
coherent and contextually relevant content. While ChatGPT- in Fig. 2, the diverse application of GAI in geoscience
like models are commonly used for tasks like generating provides unique opportunities for characterizing the Earth
text, answering questions, and providing recommendations, systems. Data collected from different resources plays a
their application in geophysical engineering is a relatively critical role in building GAI systems. Several non-exhaustive
unexplored topic with great potential. On the other hand, the potentials of generative AI in geoscience include:
integration of AI in geoscience is not without challenges.
One of the main hurdles is the lack of unbiased, high-quality, Reservoir Engineering: Generative AI can be used in reser-
labeled data for training AI models. Geophysical data is often voir engineering to examine fluid flow, well performance, and
unstructured and heterogeneous, making it difficult to use for production data.
 3

standards for carbon emissions by developing cleaner produc-

tion processes [48].

III. OVERVIEW OF G ENERATIVE AI M ODELS IN

G EOSCIENCE
In the past, machine learning and deep learning models
have been widely used for solving geo and Earth science
problems [49]. GAN-based GAI models and LLM-based tools
have found interesting applications in the field of geoscience
for various functionalities. To incorporate geoscience domain
knowledge into AI systems, physics-informed neural networks
(PINNs)-based methodologies combine robust domain knowl-
edge of the dynamics along with deep learning frameworks
for modeling and simulation. A pictorial view of these GAI
methods for geoscience is portrayed in Fig. 3.

A. GAN-based Architectures
GANs [4] emerged as a fundamentally novel approach for
solving synthetic data generation problems in computer vision
Fig. 2. Some potential application domains of generative AI in geoscience. applications. GANs framework establishes an adversarial min-
max game between its generator and discriminator networks.
During training, the generator module endeavors to produce
Reservoir Characterization and Modeling: LLMs could realistic synthetic data by sampling from a probability dis-
assist in data interpretation. By analyzing vast amounts of tribution, such as Gaussian, and utilizes a neural network as
geological and geophysical data, the models can offer valuable a universal approximator function to map the data domain.
insights to help engineers make decisions. Conversely, the discriminator module is trained to differentiate
Production Optimization: LLMs can assist in providing real- synthetic data from real data samples, creating an adversarial
time analysis of production data and suggesting adjustments training dynamic between the networks. In the realm of
to extraction rates, pressure levels, and other operational geoscience, GANs have become indispensable for diverse ap-
parameters. plications, including generating realistic seismic images [47],
Drilling: LLMs could be used to analyze geological data and simulating geological structures [46], augmenting remote sens-
predict drilling risks, which are important for good planning. ing data [50], designing road networks in futuristic cities [51],
Moreover, the models can also help determine optimal drilling and reconstructing 3-dimensional porous media [45], among
parameters and provide insights into drilling fluid properties others. This section discusses several GAN-based models and
and their effects on drilling efficiency. Additionally, GPT- their applicability in geoscience contexts.
based models can help monitor drilling operations, identify SeismoGen for Earth Data. Introduced by Wang et
possible problems, and suggest solutions. al. [47], the SeismoGen framework seeks to enhance the
Production: LLMs can analyze production data and iden- generalization capabilities of earthquake detection methods
tify potential production problems aiming for optimal perfor- based on machine learning and deep learning by supple-
mances. menting real Earth data with synthetic seismic waveforms.
Geomechanics: By analyzing geophysical data, GPT-based This GAN-based model demonstrates the capacity to generate
models can provide useful information about rock strength and three-component waveforms with multiple labels. Utilizing the
permeability, for instance. It can also help predict potential conditional GAN framework [52], the SeismoGen architecture
geomechanical risks concerning subsidence, fault reactivation, enables the incorporation of domain-specific knowledge into
and induced seismicity. synthetic data. The generator in the SeismoGen framework
Code Generation: It is well known that GAI models are is composed of a four-layered convolutional neural network,
good at code generation as they can automatically generate designed to generate synthetic waveform data and correspond-
programming scripts from high-level descriptions given by ing labels using a Gaussian noise vector and input labels.
a user (“programmers”). This is partially possible thanks to The discriminator network, on the other hand, consists of
the advanced natural language processing capabilities of LLM two sequential modules, namely the feature extractor and
models. However, the automatically produced code may not the sample critic. The former module of the network learns
be always fully understandable. a feature representation from input seismic signals and the
Carbon Emissions Reduction and Transition to Renewable latter one provides critiques based on the learned features.
Energies: As recently discussed at the 2023 United Nations To optimize the minimax game between the generator and
Climate Change Conference (COP28) in the United Arab discriminator, and facilitate a stable learning process, the value
Emirates, generative AI can indeed help fulfill the compliance function of the Wasserstein GAN [53] is employed. Thus,
 4

Fig. 3. Popularly-used Generative AI models for geoscience applications. These deep learning (DL) frameworks learn features automatically from the data by
extracting information from it. GANs work as an unsupervised framework, whereas PINNs are hybrid physics+DL models for modeling system dynamics and
solving differential equations. Foundation models mainly focus on in-context learning by using a pre-trained architecture. These models are widely applicable
to geoscience applications.

the generator loss function (LG ) and the discriminator loss selected Indian small and medium-sized cities and derived its
function (LD ) can be mathematically expressed as: transportation network index using human settlement indices.
This hybrid GAI model is built on open-source datasets,
LG = − Ez∼N(0,1) D(G(z)),
namely WSF2019 and OSM data [61]. In the hybridization,
LD =Ez∼N(0,1) D(G(z)) − Ex∼pR D(x) KRR was used to predict the transportation index of Indian
h i
2 cities of the future. This urban science-based GAI approach
+ λEz∼N(0,1) (∥D(G(z))∥2 − 1) ,
appeared valuable in situations where real data for devel-
where z represents the Gaussian noise vector, x represents the oping and underdevelopment notions is limited, as it can
real data from pR , and λ is the hyperparameter. The primary generate data closely resembling the actual urban morphology.
goal is to minimize the disparity between authentic seismic However, there exist several limitations of these tools for
waveforms and those synthetically generated by SeismoGen. urban science applications, such as validation of the simulated
This involves iterative optimization of LG and LD to es- data and dealing with time-varying urban patterns as human
tablish equilibrium between the generator and discriminator settlement changes over time.
networks. The conditional generation feature of SeismoGen
has proven its capability to produce highly realistic labeled GANs for Geological Facies Modeling. Geological fa-
seismic waveforms, proving valuable for waveform analysis cies modeling is an essential foundation for exploring and
and data augmentation [54]. characterizing sub-surface reservoirs. Traditional geostatisti-
GAI Tools for Urban Science. Application of GANs in cal models have been widely used within the geoscientific
urban science became a reality with the help of CityGAN [55], community for decades to generate sub-surface models. Re-
which learns architectural features of major cities and produces cent advancements in the field of GAI enabled the use of
images of futuristic buildings that do not exist. CityGAN is GANs combined with training-image-based simulation for
used to generate a “urban universe” from global urban land-use geological facies modeling [46]. Zhang et al.extended GANs
imageries. Another variant of CityGAN, Metropolitan GAN to 3D geological modeling at the reservoir scale to gen-
(MetroGAN) [56] is based on progressively growing [57] erate a wide variety of geologically realistic facies models
least square loss function [58] and incorporates geographical constrained by well data interpretations. Their geomodeling
loss and enforces constraints based on physical geography. approach has been validated on models such as complex
The robust urban morphology simulation capabilities of the fluvial depositional systems and carbonate reservoirs. This
MetroGAN covering several commendable strengths, however, framework uses both perceptual loss and conceptual loss in
encounter difficulties in efficiently representing all intricate the loss function where contextual loss is calculated using
features of complex urban systems. RidgeGAN combines a a distance transformation which measures the dissimilarity
kernel ridge regression (KRR) [59], [60] with the generative between GAN-generated sample and facies observation at each
CityGAN to address the problem of imaging future cities well location. To incorporate uncertainty quantification within
of India [51]. RidgeGAN captured the urban morphology of GANs, a Bayesian GAN was introduced to create a facies
 5

model with an increasing computational cost [62]. Bayesian a robust Vision LLM-based remote sensing (RS) architecture,
GANs have been applied to several geological depositional exhibits exceptional performance across various RS tasks,
scenarios to capture the variability of the data. U-Net GAN including image and region captioning, visual interrogation,
framework emerged as an alternative solution for subsurface scene classification, referring detection, and visually grounded
geological facies with fragmentary measurements [63]. This conversations. This architecture employs an innovative data
study used deep convolutional GANs to produce unconditional generation pipeline, utilizing object detection datasets [72]
facies models and the U-Net to model conditional geological for image description and Vicuna-V1.5 [69] for text-based
facies. However, most of these methods do not consider the conversations. Additionally, visual question-answering and
scales of the geological features, which remains a future scene classification capabilities are integrated using respective
challenge for the geoscientific community [64]. RS datasets. Structural enhancements in GeoChat include a
task token component specifying the desired task type, such
as grounding, image level, or region-level conversation. The
B. Foundation Models in GeoScience
framework also incorporates spatial location representation for
Foundation models in geoscience possess the capabilities to precise location specification. Visual representation, generated
tackle challenges in multimodal and multidimensional datasets by pre-trained CLIP-ViT (L-14) [73], is projected onto the
with numeric, text, audio, and video inputs. GAI models em- language model space using a single-layered neural network
ploy a scientific representation of geoscience domain knowl- adaptor with GeLU activation function [74]. Language tokens
edge that enables us to engage in reasoning processes and are then processed in Vicuna-V1.5 [69], serving as a unified
provides insights for geoscientists. The key benefits of using interface for diverse Vision LLM inputs. Efficient fine-tuning
foundation models in geoscience are scalability (data size, of the architecture with desired RS knowledge is achieved
model training, and computational power do not make a using Low-Rank Adaptation (LoRA) [75], ensuring training
problem), generalizability (superior performance on new tasks efficiency while retaining essential knowledge for contextual
going beyond the training data), dynamic interactions (con- understanding within the remote-sensing reasoning framework
volutional operations and multimodality), and flexibility [41]. of GeoChat. This efficient training positions GeoChat as a
A schematic diagram representing the workflow of foundation prominent Vision LLM in remote sensing, particularly ex-
models in geoscience is presented in Fig. 4. celling in multimodal conversational capabilities with high-
K2: An LLM Model for Geoscience. K2 is one of the resolution RS imagery for grounding, region-specific, and
first open-source LLMs specialized in geoscience [65]. Large image-level queries. A simple question-and-answer conversa-
language models, such as ChatGPT, are indeed known to yield tion with GeoChat is presented in Fig. 4 for demonstration.
impressive results in general-purpose language understanding TimeGPT for Time Series Data. Time series data plays
and generation. They acquire this amazing ability by learning a pervasive role in the field of geoscience. Given the intri-
from massive amounts of training data. However, general- cate dynamics and ever-evolving nature of Earth’s systems,
purpose LLMs may not be good enough in specialized topics continuous monitoring and analysis of temporal changes be-
such as geoscience. To design a specialized LLM for geo- come imperative. The modeling and forecasting of tempo-
science, one needs to (1) collect a large amount of geoscience ral dimensions are pivotal elements in comprehending and
knowledge, (2) consider a general-purpose LLM and train it predicting trends across diverse geoscientific phenomena en-
with the geoscience corpus, and 3) fine-tune the model with compassing climatic patterns, seismic activity, ocean currents,
supervised geoscience data. This is basically the idea behind urban morphology, and an array of other intricate variables.
K2 proposed by Deng et al. [65]. More specially, the authors These geoscientific forecasts are crucial for risk assessment,
in [65] trained a general-purpose LLaMA model [66] with climate analysis, resource management, and urban planning.
5.5B tokens of geoscience literature (Wikipedia pages, geo- It predicts natural hazards, aids in sustainable resource use,
science paper’s abstracts, open-access geoscience papers from ensures public safety, supports energy production, and informs
top journals in Earth science, and other sources from Internet), policymaking. To address the need for real-time accurate
and utilized geoscience-supervised data (called GeoSignal) to forecasts, Garza et al. [76] introduced TimeGPT, the first
fine-tune the model. Experiments on a benchmark dataset, foundation model for time series forecasting. The pre-trained
called GeoBench, showed that the K2 model is able to answer TimeGPT framework is capable of generating accurate zero-
geoscience questions and is useful for research assistance and shot predictions on unseen datasets through efficient transfer
knowledge reasoning. K2 is also claimed to be able to generate learning. The foundation of the TimeGPT framework adopts
new geoscience ideas. The model was compared favorably a similar philosophy to the large language models. The ar-
against similar-size baseline models such as Galactica-6.7B chitectural design incorporates the self-attention mechanism
[67], MPT-7B [68], Vicuna-7B [69], LLaMA-7B [66] and from the Transformers model [7] for conducting zero-shot
Alpaca-7B [70]. Such domain-specific LLM is undoubtedly inference on novel time series. Within the encoder-decoder
valuable for engineers and researchers in geoscience. Transformer architecture of the TimeGPT framework, the input
GeoChat: A Vision LLM. This recent advancement em- data, consisting of historical and optional exogenous data, is
bodies a revolutionary shift in geoscience through the integra- modeled. The encoder’s attention mechanism learns diverse
tion of Large Vision Language Models (Vision LLM), reshap- properties from the inputs, which are subsequently conveyed
ing the analysis and comprehension of Earth’s intricate dynam- to the decoder for forecasting. The sequence of predictions
ics. Introduced by Kukreja et al. [71], the GeoChat framework, concludes upon reaching the user-defined forecast horizon.
 6

Fig. 4. Overview of a Generative AI model pipeline for Geoscience Applications. Multimodal self-supervised learning algorithms train various data types
(image, text, speech, numerical data) using multidimensional geoscience data models from satellite imagery, weather, earth observations, and rivers. An example
of prompt engineering using GeoChat is represented (right below) for demonstration.

Additionally, TimeGPT, coupled with conformal prediction quickly evaluate the solution of the chosen ODEs and PDEs
techniques [77], [78], produces probabilistic forecasts and at any point in the domain of interest. Thus, PINNs have
conducts anomaly detection without the need for specific been used in ML literature as a potential candidate for solving
dataset training, a significant advantage for small-sample size differential equations [85] and prediction problems in climate
geoscience datasets. However, TimeGPT still do not work science [86], [87]. A classical problem in geophysics is to
with spatio-temporal data that are omnipotent in geoscience solve the wave equation for a given number of different
domains. sources (viz. forcing terms) which has been solved using
PINNs in the recent literature [88], [89]. To illustrate how
C. Physics-Informed Neural Networks in GeoScience PINNs work, we consider a simple PDE with two independent
Scientific machine learning (SciML) and Physics-informed variables, here denoted with t, x and u(t) is the dependent
neural networks (PINNs) are flourishing research areas of variable. We are interested in f which is a generic linear or
scientific computing aiming to build algorithms that combine nonlinear function of u(t), and α is the free parameter:
purely data-driven methods with model-driven ones. Deep ∂u(t, x) ∂u(t, x)
learning methods are good at discovering hidden patterns in + = f (u (t, x; α))
∂t ∂x
large amounts of structured data for computer vision tasks We write it more formally as:
and large language modeling tasks [79]. However, during the
model training, very little is known about the ‘physics’ or PDE (u (t, x)) = 0
underlying (dynamical) systems or processes that one wants to ∂u(t, x) ∂u(t, x)
learn from. This physical knowledge is usually in the form of ⇒ + − f (u (t, x; α)) = 0
∂t ∂x
an ordinary differential equation (ODE), a partial differential IC: u(t = 0) = u0 (x)
equation (PDE), or any other equation having an analytical
BC: u(x = x0 ) = ux0 (t); u(x = x1 ) = ux1 (t)
or numerical solution. PINNs build a model to understand a
physical process in an unsupervised manner by including the where we have specified the initial conditions (IC) and bound-
ODEs and PDEs within the loss function of the neural network ary conditions (BC). Then, a simple PINN model for the
using the training stage [80]–[84]. Automatic differentiation above PDE equation can be visualized in Fig. 3, where L
(AD) is used to compute the gradient of the loss function denoting the loss function and fθ a simple feed-forward neural
and also to compute the derivatives of the output(s) of the network for forward modeling. For a detailed discussion,
network over the inputs. Once training is done, PINNs can readers can see [90]. An application of such a scenario in
 7

geoscience problems is the classical problem of traveltime with varying degrees of incompleteness, and uncertainties.
tomography where the Eikonal equation is used as the PDE While generative AI has shown immense potential in vari-
and PINNs have shown to represent an appealing solution [91]. ous domains, its resource-intensive nature may hinder real-
Knowledge-based deep learning is an example of a PINN- time use and scalability. In fact, large-scale generative AI
based structure used for modeling the El Niño dataset [86] models usually require a significant amount of computational
whereas the Van der Pol neural network is used for multi- resources and electrical power to operate, resulting in elevated
step ahead forecasting of seismic waves, temperature, and energy consumption and significant carbon emissions. This
wind speed using PINN framework [87]. Other emerging may, unfortunately, restrict their usage in many real-world
applications of PINNs in the field of geoscience are use cases applications. Future efforts are then needed to design novel and
of time-frequency domain wave equations [92], Navier-Stokes efficient AI architectures capable of generating high-quality
equation [93], [94], and inverse problems in geophysics [95], data points in real time, which is needed for constrained
[96]. platforms. GAI tools use high computation power that has
huge environmental impact and cost inefficient. ChatGPT-
IV. B ENCHMARK DATASETS AND R ESOURCES IN based models are “black-box” in nature which opens doors
G EOSCIENCE GAI for future researchers and potential applications in geoscience.
Domain knowledge in geoscience plays a vital role in building
Data types in geoscientific fields range from space and air to
GAI models. However, inadequate data for any scientific class
ocean and ground data, among many others. A non-exhaustive
may include biases in GAI models and this could be a potential
list of datasets, tools and resources for GAI in geoscience is
risk for geoscientists during decision making. As generative
given in Table I. We highlight in Table II a refined perspective
artificial intelligence models can be considered to be still in
on the multifaceted nature of these datasets.
their infancy, here are some of the key challenges that still
As depicted in Fig. 5, data from geoscience components
have to be tackled to ensure applicability in geoscience (also
brings multimodal observations that are used for model train-
see Fig. 6):
ing and inference. Generative AI models play a vital role in
analyzing data from Atmospheric Science, Planetary Science,
• High Computational Power: Generative AI models usu-
Solid Earth Science, and Marine Science [41]. Space-based
ally consist of billions of parameters. Large-scale com-
data (Planetary Science) gives an overview of the Earth’s
puting infrastructures are usually necessary to develop
surface and atmosphere. Air-based data (Atmospheric Science)
such large models. For example, diffusion models require
provide insights about space and terrestrial observations us-
millions or billions of images to train. To train with such
ing high-quality data acquisition, Ground-based data (Solid
large datasets, massive computing power (clusters with
Earth Science) is collected from geographical or geological
hundreds of TPUs/GPUs) may be needed.
occurrences. Lastly, ocean-based data (Marine Science) gives
• Generation Latency: There is often some latency present
insights into the evaluation of marine and climate conditions.
in the time it takes to generate a new sample, mainly due
Geoscience foundation models are transforming the field of
to the large scale of generative AI models. If we take as
research with profoundly significant frontiers and their ap-
an example the diffusion models [6], which are popular
plications are evident in Earth system [49]. The availability
in the generation of high-quality samples, their sampling
of geoscience data will lead to some excellent GAI models,
speed is, however, known to be slow.
especially with open-source efforts from both industries and
• Data Scarcity: Among the main challenges in using
academia.
AI in general is the lack of large datasets for training.
Generative AI models usually require a large amount
V. C HALLENGES OF U SING G ENERATIVE AI IN of high-quality and unbiased data to operate. Lots of
G EOSCIENCE research has indeed been devoted to coping with the
The use of traditional machine learning algorithms in geo- problem of data scarcity in AI. Although generative AI
science applications is quite limited due to a variety of factors. models can be used to produce synthetic data for training,
First of all, the nature of geoscience processes poses certain other strategies can also be devised for scenarios with
inherent obstacles. For example, geoscience objects generally limited data. This includes few-shot learning [114], trans-
have amorphous boundaries in time and in space, while objects fer learning [115], and domain adaptation [116], which
in other domains are more crisply defined [113]. Hence, the offer the potential of enhancing the AI performance when
nature of geoscience objects that exist in continuous spatiotem- the data is scarce.
poral fields is much more complex than that found in discrete • Explainability and Trustworthiness: Lack of trans-
spaces that machine learning algorithms typically deal with. parency and explainability are among the other barriers
Moreover, geoscience phenomena are generally multi-variate, to the ubiquitous integration of generative artificial in-
following non-linear relationships (e.g., chaotic), having non- telligence in many real-world problems. A possible step
stationary properties, and containing unusual but fascinating towards trustworthy AI is to develop explainable AI.
events. In addition to the intrinsic difficulties of geoscience Explainable AI refers to models capable of generating de-
processes, the methodologies for gathering geoscience data cisions that a human could understand and interpret. This
present additional difficulties for machine learning algorithms. would strengthen the trust. TrustLLM can be a potential
This includes the presence of samples at multiple resolutions, solution for this problem for the scientific community and
 8

TABLE I
OPEN - SOURCE DATASETS , TOOLS , AND RESOURCES FOR AI APPLICATIONS IN G EOSCIENCE DATA

Datasets & Toolbox Brief Description Link Ref.

GeoImageNet A benchmark dataset consisting of multisource natural fea- https://github.com/ASUcicilab/GeoImageNet [97]
tures (e.g., basins, bays, islands, lakes, ridges, and valleys)
for supervised machine learning and AI in Geoscience.
BigEarthNet A benchmark archive consisting of over 590k pairs of https://bigearth.net/ [98]
Sentinel-1 and Sentinel-2 image patches that were annotated
with multi-labels of the CORINE Land Cover types to support
deep learning studies in Earth remote sensing.
EarthNets An open-source platform that links to hundreds of datasets, https://earthnets.github.io/ [99]
pre-trained deep learning models, and various tasks in Earth
Observation.
Microsoft Building Footprints Microsoft Maps & Geospatial teams released open building https://www.microsoft.com/en-us/maps [100]
footprints datasets in GeoJSON format in the United States,
Canada, and Australia, as well as many countries in Africa
and South America.
ArcGIS Living Atlas A large collection of geographic information (including maps, https://livingatlas.arcgis.com/ [101]
apps, and GIS data layers) from around the globe. It also in-
cludes a set of pre-trained deep learning models for geospatial
applications such as land use classification, tree segmentation,
and building footprint extraction.
MoveBank A publicly archived platform containing over 300 datasets https://www.movebank.org/ [102]
that describe the movement behavior of 11k animals.
Geolife GPS Trajectories An open dataset contains 17,621 GPS trajectories by 182 https://www.microsoft.com/en-us/download/ [103]
users in a period of over three years with activity labels such details.aspx?id=52367
as shopping, sightseeing, dining, hiking, and cycling.
Travel Flows A multiscale dynamic origin-to-destination population flow https://github.com/GeoDS/COVID19USFlows [104]
dataset (aggregated at three geographic scales: census tract,
county, and state; updated daily and weekly) in the U.S.
during the COVID-19 pandemic.
Google Earth Engine A multi-petabyte catalog of satellite imagery and geospa- https://earthengine.google.com/ [105]
tial datasets with planetary-scale analysis capabilities and
the Earth Engine API for geo-computation and analysis is
available in JavaScript and Python, e.g., the geemap package
ArcGIS GeoAI Toolbox A ready-to-use tool for training and using machine/deep https://pro.arcgis.com/en/pro-app/latest/ [106]
learning models that perform classification and regression on tool-reference/geoai
geospatial feature layers, imagery, tabular, and text datasets.
CyberGISX An open platform for developing and sharing open educa- https://cybergisxhub.cigi.illinois.edu/ [107]
tional resources (e.g., Jupyter Notebooks) on computationally
intensive and reproducible geospatial analytics and workflows
powered by CyberGIS middleware and cyber infrastructure.

TABLE II
G EOSCIENCE DATA TYPES , THEIR DESCRIPTIONS , AND SOURCES .

Data type Data Description Geoscience Components GAI Tools

Climate and Weather This type of data comprises data on temperature, Atmospheric Science (air-based TimeGPT, PINNs
data [108] precipitation, solar radiation, wind, air, pressure, etc. data) and Marine Science (ocean-
These datasets are used for weather forecasting and based data)
understanding climate change
Time series and Climate and seismic datasets are mostly time series Planetary Science (space-based TimeGPT, PINNs, MetroGAN
geospatial data [109] data incorporating temporal observations. Geospatial data)
data are facilitated through Geographic Information
Systems (GIS), e.g., geographic coordinates, spatial
databases, etc.
Satellite and remote Satellite imagery and remote sensing data encompass Planetary Science (air-based data) TimeGPT, PINNs, CityGAN,
sensing data [110] optical, infrared, radar, and lidar datasets. These are RidgeGAN
used to understand land cover land use (LULC) clas-
sification, vegetation, and environmental changes.
Geological and geo- These datasets include seismic readings, gravity mea- Solid Earth Science and Marine Bayesian GAN, U-Net, Seis-
physical data [111] surements, sea-surface temperature, well logs, and Science moGen
geological maps. It can be used for understanding
subsurface structures and rock properties.
Environmental These datasets include air and water quality, air Atmospheric Science and Solid PINNs, Graph Neural Net-
data [111] pollution, and environmental hazards data. It can be Earth Sciences works
used for monitoring environmental conditions and
pollution control.
Text and literature Scientific literature on geoscience, research publica- Geoscience books and publications K2, GeoChat
data [112] tions, field reports, etc., serve as valuable resources
for the geoscience community. These datasets are
useful for training LLMs for providing domain
knowledge
 9

Fig. 5. Connectionism between Geoscience foundation models and their broad usage in geosciences and related technologies. (Left) Data collected from
various sources in geoscience, including space, air, ground, and ocean machinery, provide multimodal data for model training that analyzes the data using
GAI models and performs tasks in real-time that are specified by the user to support geoscience research and advancements.

geoscientists [117]. This is not only limited to privacy- AI can indeed unintentionally produce incorrect information
sensitive research problems like in medical science but due to biases in the training data. In this context, the Eu-
also to economically critical operational decisions. ropean Union has just released some future regulations on
• Fake Data Generation and Misuse: These technologies the use of Artificial Intelligence and Large Language Models.
can also be used in the context of geoscience and remote The provisional regulations indicate that foundation models
sensing for wicked cases to create fake satellite images, must comply with specific transparency obligations. Moreover,
among many others [42]. For example, fake data gener- generative models usually require a large amount of high-
ation in geoscience may result for a non-expert user of quality data to operate. Other issues concern the latency for
information and evidence of something that seldom exists generating high-quality samples and the massive computing
and hides the truth. A real-life example could be a country power that is needed to train generative models. Lack of
that is trying to hide some land surface changes in an area explainability is also among the barriers to the ubiquitous
to mask human right violations [118]. AI-generated data integration of generative artificial intelligence. So, in the end,
could replace the same places as real satellite imagery can we really trust the outcomes of generative AI models
and can be possibly misused to fake the reality of the for critical operational decisions in geoscience? This calls for
situation from the public. more responsible, explainable, and trustworthy AI models in
different fields, including geoscience.
VI. D ISCUSSION When using generative artificial intelligence, it is then
The integration of generative AI promises a paradigm shift important to be aware of the risks and the challenges the
in geoscience. As we strive for enhanced safety, efficiency, and models are facing. For instance, to mitigate the challenge
sustainability, generative AI emerges as a key differentiator. It of the lack of data in geoscience, one can impose some
can help build a sustainable, prosperous, and technologically specific constraints to “guide” the model to learn faster and
advanced research area in geoscience. It can accelerate the better during the training. This is an effective way not only
transition towards a greener, more resilient landscape and to mitigate the shortage of training data but also to increase
environmental sustainability. However, with every innovation models’ generalizability and to ensure the physical plausibility
comes skepticism. The development of generative models of the results. This is basically the idea behind PINNs that
raises serious concerns about the potential emergence of super- take into consideration prior knowledge about the physical
intelligent machines without adequate safeguards. Generative laws that govern the data in the learning process. PINNs
 10

Fig. 6. Some challenges of using generative artificial intelligence in geoscience. GAI models use high computational power, consume time during training,
lack transparency, and sometimes are misused for fake data generation. These are the potential threats for adapting GAI models for geoscientific advancements.

overcome the low data availability issue [86] in which GANs expertise. Some technical problems may necessitate a thorough
and Transformers may lack robustness [119], rendering them comprehension of the underlying phenomena best mastered by
less effective. PINNs are shown to be effective and popular in human experts.
geophysics [120]. In conclusion, AI is promised to play a pivotal role in the
future of geoscience. It is a powerful tool that can transform
the way we interpret and understand geophysical data. While
VII. C ONCLUSION
there are challenges to overcome, the potential benefits far
In this paper, we discussed the potential of generative AI outweigh the hurdles. As we continue to harness the power
and large language models for understanding and predicting of generative AI, we can look forward to a new era of
geoscience and earth system dynamics. We introduced genera- innovation and progress in geoscience. Going forward, joint
tive AI as a remarkable tool with amazing capabilities across a efforts between geoscientists, computer scientists, and the AI
broad spectrum of tasks. Then, we discussed the integration of community can make discoveries in pressing challenges like
generative AI in geoscience, arguing the relatively unexplored climate change, environmental hazards, and sustainability. This
territory with great potential. We also enumerated a non- paper is intended to provide guidance but does not claim to
exhaustive list of potential applications of generative AI in be comprehensive or conclusive in any way.
geoscience. As every innovation comes with skepticism, we
then presented some of the key challenges that still have to ACKNOWLEDGMENTS
be tackled to ensure the applicability of generative AI in The support of TotalEnergies is fully acknowledged. Ab-
geoscience. To further explain the use of generative AI in denour Hadid (Professor, Industry Chair at SCAI Center of
geoscience, a concrete example was given, describing models Abu Dhabi) is funded by TotalEnergies collaboration agree-
based on GANs for data generation in geoscience. The model ment with Sorbonne University Abu Dhabi. The authors also
is a seismic waveform synthesis technique that exploits GAN acknowledge open-source generative AI tools that have been
for seismic data generation. Moreover, several examples of used for the generations of images used in this paper. T.C.
large language models capable of understanding geoscience acknowledges Madhurima Panja of IIIT Bangalore, India for
and generating new geoscience ideas are also described. valuable insights and suggestions.
Lately, we have discussed the importance of mitigating the AI
challenges and, specifically, the lack of data in geoscience to
R EFERENCES
train the AI models by imposing specific constraints to “guide”
the model training. This is an effective way to mitigate the [1] Y. Cao, S. Li, Y. Liu, Z. Yan, Y. Dai, P. S. Yu, and L. Sun, “A
comprehensive survey of ai-generated content (aigc): A history of
shortage of training data, to increase models’ generalizability generative ai from gan to chatgpt,” arXiv preprint arXiv:2303.04226,
and to ensure the physical plausibility of the results (as done 2023.
in PINNs). Other limitations of GAI models in geoscience in- [2] S. R. Eddy, “Hidden markov models,” Current opinion in structural
biology, vol. 6, no. 3, pp. 361–365, 1996.
volve scalability, interpretability, trustworthiness, social biases, [3] L. R. Rabiner, “A tutorial on hidden markov models and selected
and fake data generation have also been highlighted. Improved applications in speech recognition,” Proceedings of the IEEE, vol. 77,
scalability and generalizability can make GAI models tackle no. 2, pp. 257–286, 1989.
[4] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley,
the challenges of Earth’s systems. S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets
The use of generative AI in geoscience holds the promise (advances in neural information processing systems)(pp. 2672–2680),”
of transforming various aspects of the industry, from reservoir Red Hook, NY Curran, 2014.
[5] J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli,
analysis and production optimization to safety protocols and “Deep unsupervised learning using nonequilibrium thermodynamics,”
collaborative decision-making. As AI technology continues in International conference on machine learning. PMLR, 2015, pp.
to advance, the integration of generative AI into geoscience 2256–2265.
[6] L. Yang, Z. Zhang, Y. Song, S. Hong, R. Xu, Y. Zhao, W. Zhang,
could then mark a significant step forward in the industry’s B. Cui, and M.-H. Yang, “Diffusion models: A comprehensive survey
pursuit of innovation and optimization. Although generative of methods and applications,” ACM Computing Surveys, vol. 56, no. 4,
AI can offer elegant solutions to many technical challenges in pp. 1–39, 2023.
[7] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N.
geoscience, it is important to note that generative AI is not a Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”
complete solution that can replace all human experience and Advances in neural information processing systems, vol. 30, 2017.
 11

[8] D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” arXiv [32] F. F. Xu, U. Alon, G. Neubig, and V. J. Hellendoorn, “A systematic
preprint arXiv:1312.6114, 2013. evaluation of large language models of code,” in Proceedings of the 6th
[9] A. Radford, J. Wu, D. Amodei, D. Amodei, J. Clark, M. Brundage, and ACM SIGPLAN International Symposium on Machine Programming,
I. Sutskever, “Better language models and their implications,” OpenAI 2022, pp. 1–10.
blog, vol. 1, no. 2, 2019. [33] C. Thapa, S. I. Jang, M. E. Ahmed, S. Camtepe, J. Pieprzyk, and
[10] W. X. Zhao, K. Zhou, J. Li, T. Tang, X. Wang, Y. Hou, Y. Min, S. Nepal, “Transformer-based language models for software vulnera-
B. Zhang, J. Zhang, Z. Dong et al., “A survey of large language bility detection,” in Proceedings of the 38th Annual Computer Security
models,” arXiv preprint arXiv:2303.18223, 2023. Applications Conference, 2022, pp. 481–496.
[11] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training [34] J. Liu, C. S. Xia, Y. Wang, and L. Zhang, “Is your code generated by
of deep bidirectional transformers for language understanding,” arXiv chatgpt really correct? rigorous evaluation of large language models
preprint arXiv:1810.04805, 2018. for code generation,” arXiv preprint arXiv:2305.01210, 2023.
[12] A. Radford, K. Narasimhan, T. Salimans, I. Sutskever et al., Improving [35] C. Fieberg, L. Hornuf, and D. Streich, “Using gpt-4 for financial
language understanding by generative pre-training. OpenAI, 2018. advice,” Available at SSRN 4488891, 2023.
[13] P. P. Ray, “Chatgpt: A comprehensive review on background, appli- [36] A. Zaremba and E. Demir, “Chatgpt: Unlocking the future of nlp in
cations, key challenges, bias, ethics, limitations and future scope,” finance,” Available at SSRN 4323643, 2023.
Internet of Things and Cyber-Physical Systems, 2023. [37] G. Son, H. Jung, M. Hahm, K. Na, and S. Jin, “Beyond classification:
[14] K. Singhal, S. Azizi, T. Tu, S. S. Mahdavi, J. Wei, H. W. Chung, Financial reasoning in state-of-the-art language models,” arXiv preprint
N. Scales, A. Tanwani, H. Cole-Lewis, S. Pfohl et al., “Large language arXiv:2305.01505, 2023.
models encode clinical knowledge,” Nature, vol. 620, no. 7972, pp. [38] S. Wu, O. Irsoy, S. Lu, V. Dabravolski, M. Dredze, S. Gehrmann,
172–180, 2023. P. Kambadur, D. Rosenberg, and G. Mann, “Bloomberggpt: A large
[15] K. Singhal, T. Tu, J. Gottweis, R. Sayres, E. Wulczyn, L. Hou, language model for finance,” arXiv preprint arXiv:2303.17564, 2023.
K. Clark, S. Pfohl, H. Cole-Lewis, D. Neal et al., “Towards expert- [39] D. Araci, “Finbert: Financial sentiment analysis with pre-trained lan-
level medical question answering with large language models,” arXiv guage models,” arXiv preprint arXiv:1908.10063, 2019.
preprint arXiv:2305.09617, 2023. [40] X. Zhang and Q. Yang, “Xuanyuan 2.0: A large chinese financial chat
[16] V. Liévin, C. E. Hother, and O. Winther, “Can large language models model with hundreds of billions parameters,” in Proceedings of the
reason about medical questions?” arXiv preprint arXiv:2207.08143, 32nd ACM International Conference on Information and Knowledge
2022. Management, 2023, pp. 4435–4439.
[17] H. Nori, N. King, S. M. McKinney, D. Carignan, and E. Horvitz, [41] H. Zhang and J.-J. Xu, “When geoscience meets foundation mod-
“Capabilities of gpt-4 on medical challenge problems,” arXiv preprint els: Towards general geoscience artificial intelligence system,” arXiv
arXiv:2303.13375, 2023. preprint arXiv:2309.06799, 2023.
[18] P. Sharma, K. Thapa, P. Dhakal, M. D. Upadhaya, S. Adhikari, and [42] N. Patel, “Generative artificial intelligence and remote sensing: A
S. R. Khanal, “Performance of chatgpt on usmle: Unlocking the perspective on the past and the future [perspectives],” IEEE Geoscience
potential of large language models for ai-assisted medical education,” and Remote Sensing Magazine, vol. 11, no. 2, pp. 86–100, 2023.
arXiv preprint arXiv:2307.00112, 2023.
[43] A. Karpatne, I. Ebert-Uphoff, S. Ravela, H. A. Babaie, and V. Kumar,
[19] F. Antaki, S. Touma, D. Milad, J. El-Khoury, and R. Duval, “Evaluating
“Machine learning for the geosciences: Challenges and opportunities,”
the performance of chatgpt in ophthalmology: An analysis of its
IEEE Transactions on Knowledge and Data Engineering, vol. 31, no. 8,
successes and shortcomings,” Ophthalmology Science, p. 100324, 2023.
pp. 1544–1554, 2018.
[20] D. M. Levine, R. Tuwani, B. Kompa, A. Varma, S. G. Finlayson,
[44] K. J. Bergen, P. A. Johnson, M. V. de Hoop, and G. C. Beroza,
A. Mehrotra, and A. Beam, “The diagnostic and triage accuracy of
“Machine learning for data-driven discovery in solid earth geoscience,”
the gpt-3 artificial intelligence model,” medRxiv, pp. 2023–01, 2023.
Science, vol. 363, no. 6433, p. eaau0323, 2019.
[21] A. Tack and C. Piech, “The ai teacher test: Measuring the pedagogical
[45] L. Mosser, O. Dubrule, and M. J. Blunt, “Reconstruction of three-
ability of blender and gpt-3 in educational dialogues,” arXiv preprint
dimensional porous media using generative adversarial neural net-
arXiv:2205.07540, 2022.
works,” Physical Review E, vol. 96, no. 4, p. 043309, 2017.
[22] R. E. Wang and D. Demszky, “Is chatgpt a good teacher coach?
measuring zero-shot performance for scoring and providing actionable [46] T.-F. Zhang, P. Tilke, E. Dupont, L.-C. Zhu, L. Liang, and W. Bailey,
insights on classroom instruction,” arXiv preprint arXiv:2306.03090, “Generating geologically realistic 3d reservoir facies models using
2023. deep learning of sedimentary architecture with generative adversarial
[23] Z. A. Pardos and S. Bhandari, “Learning gain differences between networks,” Petroleum Science, vol. 16, pp. 541–549, 2019.
chatgpt and human tutor generated algebra hints,” arXiv preprint [47] T. Wang, D. Trugman, and Y. Lin, “Seismogen: Seismic waveform
arXiv:2302.06871, 2023. synthesis using gan with application to seismic data augmentation,”
[24] W. Dai, J. Lin, H. Jin, T. Li, Y.-S. Tsai, D. Gašević, and G. Chen, Journal of Geophysical Research: Solid Earth, vol. 126, no. 4, p.
“Can large language models provide feedback to students? a case e2020JB020077, 2021.
study on chatgpt,” in 2023 IEEE International Conference on Advanced [48] M. Romanello, S. Whitmee, E. Mulcahy, and A. Costello, “Further
Learning Technologies (ICALT). IEEE, 2023, pp. 323–325. delays in tackling greenhouse gas emissions at cop28 will be an act of
[25] J. Savelka, K. D. Ashley, M. A. Gray, H. Westermann, and H. Xu, negligence,” The Lancet, vol. 402, no. 10417, pp. 2055–2057, 2023.
“Explaining legal concepts with augmented large language models (gpt- [49] G. Artificial, Handbook of Geospatial Artificial Intelligence. Rout-
4),” arXiv preprint arXiv:2306.09525, 2023. ledge, 2023.
[26] I. Chalkidis, M. Fergadiotis, P. Malakasiotis, N. Aletras, and I. An- [50] N. Lv, H. Ma, C. Chen, Q. Pei, Y. Zhou, F. Xiao, and J. Li, “Remote
droutsopoulos, “Legal-bert: The muppets straight out of law school,” sensing data augmentation through adversarial training,” IEEE Journal
arXiv preprint arXiv:2010.02559, 2020. of Selected Topics in Applied Earth Observations and Remote Sensing,
[27] A. Blair-Stanek, N. Holzenberger, and B. Van Durme, “Can gpt-3 vol. 14, pp. 9318–9333, 2021.
perform statutory reasoning?” arXiv preprint arXiv:2302.06100, 2023. [51] R. Thottolil, U. Kumar, and T. Chakraborty, “Prediction of transporta-
[28] F. Yu, L. Quartey, and F. Schilder, “Legal prompting: Teaching a lan- tion index for urban patterns in small and medium-sized indian cities
guage model to think like a lawyer,” arXiv preprint arXiv:2212.01326, using hybrid ridgegan model,” arXiv preprint arXiv:2306.05951, 2023.
2022. [52] M. Mirza and S. Osindero, “Conditional generative adversarial nets,”
[29] S. I. Ross, F. Martinez, S. Houde, M. Muller, and J. D. Weisz, arXiv preprint arXiv:1411.1784, 2014.
“The programmer’s assistant: Conversational interaction with a large [53] M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein generative
language model for software development,” in Proceedings of the 28th adversarial networks,” in International conference on machine learning.
International Conference on Intelligent User Interfaces, 2023, pp. 491– PMLR, 2017, pp. 214–223.
514. [54] W. Li and J. Wang, “Residual learning of cycle-gan for seismic data
[30] G. Sandoval, H. Pearce, T. Nys, R. Karri, S. Garg, and B. Dolan-Gavitt, denoising,” Ieee Access, vol. 9, pp. 11 585–11 597, 2021.
“Lost at c: A user study on the security implications of large language [55] A. Albert, E. Strano, J. Kaur, and M. C. González, “Modeling urban-
model code assistants,” arXiv preprint arXiv:2208.09727, 2023. ization patterns with generative adversarial networks,” IGARSS 2018 -
[31] J. Leinonen, P. Denny, S. MacNeil, S. Sarsa, S. Bernstein, J. Kim, 2018 IEEE International Geoscience and Remote Sensing Symposium,
A. Tran, and A. Hellas, “Comparing code explanations created by pp. 2095–2098, 2018.
students and large language models,” arXiv preprint arXiv:2304.03938, [56] W. Zhang, Y. Ma, D. Zhu, L. Dong, and Y. Liu, “Metrogan: Simulating
2023. urban morphology with generative adversarial network,” in Proceedings
 12

of the 28th ACM SIGKDD Conference on Knowledge Discovery and [81] M. Rasht-Behesht, C. Huber, K. Shukla, and G. E. Karniadakis,
Data Mining, 2022, pp. 2482–2492. “Physics-informed neural networks (pinns) for wave propagation and
[57] T. Karras, T. Aila, S. Laine, and J. Lehtinen, “Progressive growing full waveform inversions,” Journal of Geophysical Research: Solid
of gans for improved quality, stability, and variation,” arXiv preprint Earth, vol. 127, no. 5, p. e2021JB023120, 2022.
arXiv:1710.10196, 2017. [82] L. Yang, X. Meng, and G. E. Karniadakis, “B-pinns: Bayesian physics-
[58] X. Mao, Q. Li, H. Xie, R. Y. Lau, Z. Wang, and S. Paul Smolley, “Least informed neural networks for forward and inverse pde problems with
squares generative adversarial networks,” in Proceedings of the IEEE noisy data,” Journal of Computational Physics, vol. 425, p. 109913,
international conference on computer vision, 2017, pp. 2794–2802. 2021.
[59] V. Vovk, “Kernel ridge regression,” in Empirical Inference: Festschrift [83] X. Jin, S. Cai, H. Li, and G. E. Karniadakis, “Nsfnets (navier-stokes
in Honor of Vladimir N. Vapnik. Springer, 2013, pp. 105–116. flow nets): Physics-informed neural networks for the incompressible
[60] K. P. Murphy, Machine learning: a probabilistic perspective. MIT navier-stokes equations,” Journal of Computational Physics, vol. 426,
press, 2012. p. 109951, 2021.
[61] M. Marconcini, A. Metz-Marconcini, S. Üreyen, D. Palacios-Lopez, [84] Z. Mao, A. D. Jagtap, and G. E. Karniadakis, “Physics-informed
W. Hanke, F. Bachofer, J. Zeidler, T. Esch, N. Gorelick, A. Kakarla neural networks for high-speed flows,” Computer Methods in Applied
et al., “Outlining where humans live, the world settlement footprint Mechanics and Engineering, vol. 360, p. 112789, 2020.
2015,” Scientific Data, vol. 7, no. 1, p. 242, 2020. [85] L. Lu, X. Meng, Z. Mao, and G. E. Karniadakis, “Deepxde: A
[62] R. Feng, D. Grana, T. Mukerji, and K. Mosegaard, “Application of deep learning library for solving differential equations,” SIAM review,
bayesian generative adversarial networks to geological facies model- vol. 63, no. 1, pp. 208–228, 2021.
ing,” Mathematical Geosciences, vol. 54, no. 5, pp. 831–855, 2022. [86] Z. Elabid, T. Chakraborty, and A. Hadid, “Knowledge-based deep
[63] C. Zhang, X. Song, and L. Azevedo, “U-net generative adversarial learning for modeling chaotic systems,” in 2022 21st IEEE Interna-
network for subsurface facies modeling,” Computational Geosciences, tional Conference on Machine Learning and Applications (ICMLA).
vol. 25, pp. 553–573, 2021. IEEE, 2022, pp. 1203–1209.
[64] S. Song, T. Mukerji, and J. Hou, “Geological facies modeling based [87] A. Dutta, M. Panja, U. Kumar, C. Hens, and T. Chakraborty, “Van
on progressive growing of generative adversarial networks (gans),” der pol-informed neural networks for multi-step-ahead forecasting of
Computational Geosciences, vol. 25, pp. 1251–1273, 2021. extreme climatic events,” in NeurIPS 2023 AI for Science Workshop,
[65] C. Deng, T. Zhang, Z. He, Q. Chen, Y. Shi, L. Zhou, L. Fu, W. Zhang, 2023.
X. Wang, C. Zhou, Z. Lin, and J. He, “Learning a foundation language [88] B. Moseley, A. Markham, and T. Nissen-Meyer, “Solving the
model for geoscience knowledge understanding and utilization,” in The wave equation with physics-informed deep learning,” arXiv preprint
17th ACM International Conference on Web Search and Data Mining, arXiv:2006.11894, 2020.
03 2024. [89] J. Pu, J. Li, and Y. Chen, “Solving localized wave solutions of the
[66] H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, derivative nonlinear schrödinger equation using an improved pinn
T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, method,” Nonlinear Dynamics, vol. 105, pp. 1723–1739, 2021.
A. Joulin, E. Grave, and G. Lample, “Llama: Open and efficient [90] G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang,
foundation language models,” 2023. and L. Yang, “Physics-informed machine learning,” Nature Reviews
[67] R. Taylor, M. Kardas, G. Cucurull, T. Scialom, A. Hartshorn, E. Sar- Physics, vol. 3, no. 6, pp. 422–440, 2021.
avia, A. Poulton, V. Kerkez, and R. Stojnic, “Galactica: A large
[91] U. bin Waheed, E. Haghighat, T. Alkhalifah, C. Song, and Q. Hao,
language model for science,” 2022.
“Pinneik: Eikonal solution using physics-informed neural networks,”
[68] M. Team et al., “Introducing mpt-7b: a new standard for open-source,
Computers & Geosciences, vol. 155, p. 104833, 2021.
commercially usable llms,” 2023.
[92] S. Karimpouli and P. Tahmasebi, “Physics informed machine learning:
[69] W.-L. Chiang, Z. Li, Z. Lin, Y. Sheng, Z. Wu, H. Zhang, L. Zheng,
Seismic wave equation,” Geoscience Frontiers, vol. 11, no. 6, pp. 1993–
S. Zhuang, Y. Zhuang, J. E. Gonzalez et al., “Vicuna: An open-
2001, 2020.
source chatbot impressing gpt-4 with 90%* chatgpt quality,” See
https://vicuna. lmsys. org (accessed 14 April 2023), 2023. [93] R. Ranade, C. Hill, and J. Pathak, “Discretizationnet: A machine-
[70] R. Taori, I. Gulrajani, T. Zhang, Y. Dubois, X. Li, C. Guestrin, P. Liang, learning based solver for navier–stokes equations using finite volume
and T. B. Hashimoto, “Stanford alpaca: An instruction-following llama discretization,” Computer Methods in Applied Mechanics and Engi-
model,” 2023. neering, vol. 378, p. 113722, 2021.
[71] K. Kuckreja, M. S. Danish, M. Naseer, A. Das, S. Khan, and F. S. [94] S. Hu, M. Liu, S. Zhang, S. Dong, and R. Zheng, “Physics-informed
Khan, “Geochat: Grounded large vision-language model for remote neural network combined with characteristic-based split for solving
sensing,” arXiv preprint arXiv:2311.15826, 2023. navier–stokes equations,” Engineering Applications of Artificial Intel-
[72] D. Wang, J. Zhang, B. Du, M. Xu, L. Liu, D. Tao, and L. Zhang, ligence, vol. 128, p. 107453, 2024.
“Samrs: Scaling-up remote sensing segmentation dataset with segment [95] Y. Zhang, X. Zhu, and J. Gao, “Seismic inversion based on acoustic
anything model,” in Thirty-seventh Conference on Neural Information wave equations using physics-informed neural network,” IEEE trans-
Processing Systems Datasets and Benchmarks Track, 2023. actions on geoscience and remote sensing, vol. 61, pp. 1–11, 2023.
[73] Y. Tay, M. C. Phan, L. A. Tuan, and S. C. Hui, “Learning to rank [96] I. Depina, S. Jain, S. Mar Valsson, and H. Gotovac, “Application of
question answer pairs with holographic dual lstm architecture,” in physics-informed neural networks to inverse problems in unsaturated
Proceedings of the 40th international ACM SIGIR conference on groundwater flow,” Georisk: Assessment and Management of Risk for
research and development in information retrieval, 2017, pp. 695–704. Engineered Systems and Geohazards, vol. 16, no. 1, pp. 21–36, 2022.
[74] D. Hendrycks and K. Gimpel, “Gaussian error linear units (gelus),” [97] W. Li, S. Wang, S. T. Arundel, and C.-Y. Hsu, “Geoimagenet: a multi-
arXiv preprint arXiv:1606.08415, 2016. source natural feature benchmark dataset for geoai and supervised
[75] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, machine learning,” GeoInformatica, vol. 27, no. 3, pp. 619–640, 2023.
and W. Chen, “Lora: Low-rank adaptation of large language models,” [98] G. Sumbul, M. Charfuelan, B. Demir, and V. Markl, “Bigearthnet: A
arXiv preprint arXiv:2106.09685, 2021. large-scale benchmark archive for remote sensing image understand-
[76] A. Garza and M. Mergenthaler-Canseco, “Timegpt-1,” arXiv preprint ing,” in IGARSS 2019-2019 IEEE International Geoscience and Remote
arXiv:2310.03589, 2023. Sensing Symposium. IEEE, 2019, pp. 5901–5904.
[77] V. Vovk, A. Gammerman, and G. Shafer, “Conformal prediction,” [99] Z. Xiong, F. Zhang, Y. Wang, Y. Shi, and X. X. Zhu, “Earthnets:
Algorithmic learning in a random world, pp. 17–51, 2005. Empowering ai in earth observation,” arXiv preprint arXiv:2210.04936,
[78] V. Vovk, J. Shen, V. Manokhin, and M.-g. Xie, “Nonparametric 2022.
predictive distributions based on conformal prediction,” in Conformal [100] M. P. Heris, N. L. Foks, K. J. Bagstad, A. Troy, and Z. H. Ancona,
and probabilistic prediction and applications. PMLR, 2017, pp. 82– “A rasterized building footprint dataset for the united states,” Scientific
102. data, vol. 7, no. 1, p. 207, 2020.
[79] I. Goodfellow, Y. Bengio, and A. Courville, Deep learning. MIT [101] H. West and M. Horswell, “Gis has changed! exploring the potential of
press, 2016. arcgis online,” Teaching Geography, vol. 43, no. 1, pp. 22–24, 2018.
[80] M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed [102] B. Kranstauber, A. Cameron, R. Weinzerl, T. Fountain, S. Tilak,
neural networks: A deep learning framework for solving forward and M. Wikelski, and R. Kays, “The movebank data model for animal
inverse problems involving nonlinear partial differential equations,” tracking,” Environmental Modelling & Software, vol. 26, no. 6, pp.
Journal of Computational physics, vol. 378, pp. 686–707, 2019. 834–835, 2011.
 13

[103] Y. Zheng, X. Xie, W.-Y. Ma et al., “Geolife: A collaborative social

networking service among user, location and trajectory.” IEEE Data
Eng. Bull., vol. 33, no. 2, pp. 32–39, 2010.
[104] Y. Kang, S. Gao, Y. Liang, M. Li, J. Rao, and J. Kruse, “Multiscale
dynamic human mobility flow dataset in the us during the covid-19
epidemic,” Scientific data, vol. 7, no. 1, p. 390, 2020.
[105] Q. Wu, “geemap: A python package for interactive mapping with
google earth engine,” Journal of Open Source Software, vol. 5, no. 51,
p. 2305, 2020.
[106] J. Fleming, S. W. Marvel, S. Supak, A. A. Motsinger-Reif, and D. M.
Reif, “Toxpi* gis toolkit: Creating, viewing, and sharing integrative
visualizations for geospatial data using arcgis,” Journal of Exposure
Science & Environmental Epidemiology, vol. 32, no. 6, pp. 900–907,
2022.
[107] S. Wang, L. Anselin, B. Bhaduri, C. Crosby, M. F. Goodchild, Y. Liu,
and T. L. Nyerges, “Cybergis software: a synthetic review and inte-
gration roadmap,” International Journal of Geographical Information
Science, vol. 27, no. 11, pp. 2122–2145, 2013.
[108] M. G. De Vos, W. Hazeleger, D. Bari, J. Behrens, S. Bendoukha,
I. Garcia-Marti, R. van Haren, S. E. Haupt, R. Hut, F. Jansson et al.,
“Open weather and climate science in the digital era,” Geoscience
Communication, vol. 3, no. 2, pp. 191–201, 2020.
[109] A. Lewis, S. Oliver, L. Lymburner, B. Evans, L. Wyborn, N. Mueller,
G. Raevksi, J. Hooke, R. Woodcock, J. Sixsmith et al., “The australian
geoscience data cube—foundations and lessons learned,” Remote Sens-
ing of Environment, vol. 202, pp. 276–292, 2017.
[110] R. R. Navalgund, V. Jayaraman, and P. Roy, “Remote sensing applica-
tions: An overview,” current science, pp. 1747–1766, 2007.
[111] T. S. Bressan, M. K. de Souza, T. J. Girelli, and F. C. Junior, “Eval-
uation of machine learning methods for lithology classification using
geophysical data,” Computers & Geosciences, vol. 139, p. 104475,
2020.
[112] S. Zhang, H. Xu, Y. Jia, Y. Wen, D. Wang, L. Fu, X. Wang, and
C. Zhou, “Geodeepshovel: A platform for building scientific database
from geoscience literature with ai assistance,” Geoscience Data Jour-
nal, vol. 10, no. 4, pp. 519–537, 2023.
[113] T. Chakraborty, S. M. Naik, M. Panja, B. Manvitha et al., “Ten years
of generative adversarial nets (gans): A survey of the state-of-the-art,”
arXiv preprint arXiv:2308.16316, 2023.
[114] Y. Song, T. Wang, P. Cai, S. K. Mondal, and J. P. Sahoo, “A
comprehensive survey of few-shot learning: Evolution, applications,
challenges, and opportunities,” ACM Computing Surveys, 2023.
[115] F. Zhuang, Z. Qi, K. Duan, D. Xi, Y. Zhu, H. Zhu, H. Xiong, and
Q. He, “A comprehensive survey on transfer learning,” Proceedings of
the IEEE, vol. 109, no. 1, pp. 43–76, 2020.
[116] Z. Yu, J. Li, Z. Du, L. Zhu, and H. T. Shen, “A comprehensive survey
on source-free domain adaptation,” arXiv preprint arXiv:2302.11803,
2023.
[117] L. e. a. Sun, “Trustllm: Trustworthiness in large language models,”
arXiv preprint arXiv:2401.05561, 2024.
[118] R. Rodrigues, “Legal and human rights issues of ai: Gaps, challenges
and vulnerabilities,” Journal of Responsible Technology, vol. 4, p.
100005, 2020.
[119] L. Sasal, T. Chakraborty, and A. Hadid, “W-transformers: A wavelet-
based transformer framework for univariate time series forecasting,” in
2022 21st IEEE International Conference on Machine Learning and
Applications (ICMLA). IEEE, 2022, pp. 671–676.
[120] Z. Hao, S. Liu, Y. Zhang, C. Ying, Y. Feng, H. Su, and J. Zhu,
“Physics-informed machine learning: A survey on problems, methods
and applications,” arXiv preprint arXiv:2211.08064, 2022.