Chemical Reactions Engineering Innovative Project

Chemical Reactions Engineering
Innovative project
Use of AI in
Chemical
Kinetics
SUBMITTED BY SIDDHANT VIJAY
(2K20/CH/65)
AND
MEHUL VERMA
Submitted to Respected Raminder Kaur Ma'am (2K20/CH/45)
How AI can help in Chemical Kinetics
Machine Learning and deep learning
Machine learning (ML) is a subset of
artificial intelligence (AI), which
provides systems with the ability to
automatically learn and improve the
expected response in terms of
modeling, based on experimental,
training data whereas Deep learning
is a subset of machine learning that
uses neural networks to analyze
various factors in a similar way to the
human learning process.
The key idea of ML is that it is possible
to create algorithms that execute the
learning process using training data
and create predictions and
generalizations based on it
Why do we need to use AI in Chemical Kinetics
Taking into account the fact that classical kinetic models are sufficient, it would seem that there is
no greater need to model the paths and rates of a chemical reaction using other methods
However, classical models, even those containing physicochemical interpretable constants of rate
or equilibrium constants, provide opportunities to perform wide generalization. When deep
machine learning techniques are used, additional possibilities of creating more general models,
capable of predicting the paths of reactions for which they were not originally trained, open up.
the ability to create generalizations, to find distant relationships, and to bind apparently distant
processes and phenomena makes it possible to create a single neural network, using deep learning,
that would not only be able to give an answer about the reactions it learned in the course of
training but also about those it did not know originally.
An additional benefit of using artificial neural networks is that it is a kind of data storage,
understood as knowledge of the kinetics of various chemical reactions compressed in the form of
weights on the dendrites of neurons of the artificial network.
Kinetic Models in Chemical Reactions
One of the central elements of the

chemical reaction nature is how
fast it runs. The research on the
kinetics of chemical processes has
been conducted for a long time and
in classical approaches addresses
the problem in the form of
appropriate mathematical
equations. Kinetics, in other words,
the rate of chemical reaction, is
typically described by well-known
models
Neural network training with Arrhenius kinetics for
equilibrium reactions
The method presented in fig 3 reflects the

calculation of the source term in the form
realized by means of an artificial neural
network, which allows to interpret the
equations as a series of connected neural
layers. The structure of neural network
reflects the process by means of process
variables, namely the concentrations and
temperatures. The neural network output
on the other hand is defined as the
resultant source terms set
Deep learning based on quantum
modeling of chemical reactions paths
Data on the activation energy of chemical reactions

are difficult to access due to the complexity of the
implementation of experimental studies, especially
in the case of reactions that are characterized by
high rates. Prediction of activation energy is useful
for computer-aided generation of reaction
mechanisms and planning of organic synthesis. The
authors propose a deep learning model for
prediction of activation energy of selected reactions
and geometry of product molecules. The authors
have developed a deep learning model to predict
activation energies for a wide range of reaction
types, which does not depend on any additional
input data and requires only graphical
representation of the reagents. Such a model
represents a deep learning neural network built to
estimate kinetics by automatically generating the
reaction mechanism and allows for the quantitative
ranking of candidates for possible reactions.
Deep learning based on quantum
modeling of chemical reactions paths
Methodology
The authors use the directed message passing neural network (D-
MPNN) method to predict molecular properties. The D-MPNN maintains
two representations for information placed on the bond between the
atoms v andw: one of the atoms v to (mtvw) and one of the atoms w to
(mtwv). Consequently, instead of aggregating information from
adjacent atoms, D-MPNN aggregates information from adjacent bonds.
The information about each bond is updated based on all other, that is,
not belonging to a given bonded atom, incoming information about
bonds. Because of this structure, with the information located on the
bonds and the distinction between the two directions of the bond
messages, the D-MPNN has more control over the information flow
through the molecule and therefore can build more informative
molecular representations. Modeling of chemical reactions kinetics 15
The D-MPNN constructs the learned representation of the molecule by
transferring information between the elements of the chart using
information related to atomic bonds. Binding information contains
characteristics of the atom, such as atomic number, and the initial
properties of the bond, using a neural network. The neural network
iteratively updates the geometry of the molecules by using information
from adjacent bonds. Since the mapping of atoms between substrate
and product is known, the difference between all atoms of products
and substrates is calculated to obtain socalled fingerprints of atomic
differences. In the next step each fingerprint is used by the same neural
network to create a reaction path encoding. The final stage of the
process is the reading phase, in which the network learns the linear
combination of elements in the encoding of the reaction to obtain the
estimated activation energy
Enzymatic Reaction Kinetics
The study investigated the usefulness of ANN to assess the rate of enzymatic reactions. The study was
conducted on the basis of a model reaction, the enzymatic hydrolysis of maltose, catalyzed by the
enzyme amyloglucosidase. The influence of substrate (maltose) and product (glucose) concentration on
the rate of enzymatic reaction was studied. The reaction taken into account reads

Amyloglucosidase
Maltose + H2O 2Glucose
The artificial neural network was built with The Neural Network Toolbox of MATLAB
software. The feedforward neural network was formed, which consists of three neurons
in the input layer, four neurons in the hidden layer and one neuron in the output layer.
Neural network inputs are maltose concentration, glucose concentration and initial
glucose concentration, and system output is the rate of reaction. The network was
trained until the square mean value between the training set values and the outputs was
less than 1·10−4
Benefits of using ANN in Enzymatic Reactions
First of all, ANN estimates the reaction rate without using any kinetic model equation
Secondly, estimating the reaction rate without the use of a kinetic model eliminates the
risk resulting from the wrong choice of the kinetic model
Thirdly, the use of ANN requires smaller amount of experimental data than
conventional kinetic model with similar accuracy. This is due to the general feature of
neural networks, i.e., the ability to generalize.
Fourthly, when using ANN, there is no need to derive kinetic model. As a result, the
classical kinetic model does not impose any limitations on the actual experimental
analysis of the kinetics of the reaction system. On the other hand ANN does not provide
any explanation on physicochemical nature of the process.
Hydrogen Oxidation Kinetics using ANN
How?
In the paper authors use a deep learning of the ANN feedforward type to model the H2 oxidation reaction.
The machine learning algorithm was adapted to this case, in such a way which enabled to act as a universal
approximation, improving in areas of high nonl inearity, noise and background signal bias. This approach
allows for flexibility in both network architecture and loss functions, as well as for scalability on modern
architectures. The primary goal of implementing a deep neural network is to approximate functions f:x → y
based on input x and output y. After selecting function f, the network weights are adjusted by a process
called stochastic gradient descent in which the model:
– iteratively performs predictions on randomly selected sets of training examples,

– determines the quality of the prediction using the L loss function,
– updates parameter weights in a process called back propagation.
SUMMARY
Although quite a few of the cases presented above claim to use deep learning techniques it seems to
be a bit of exaggeration, at least at the neural network cognitive level. Deep neural network refers
more to the cognitive capabilities of the artificial intelligence than to the way it is assembled.
Although it is obvious that deep learning demands that the network construction is made of higher

number of layers and neurons. Each case presented although was successful to show the potential of
trained neural network prediction in typical engineering and design calculations.
Future perspective for modeling the chemical reactions paths and rates by artificial intelligence
systems is promising. Artificial neural network might be understood in simple picture as a huge set
of regression variables, which in fact are weights on the neurons of the neural layers, to be adjusted
during the training process. Thus the ability of neural network to generate comprehensive and
generalized responses to given problem allows to create predictions even for the chemical and
physical systems it was not trained. Such behavior resembles human brain functions in its
fundamental cognitive roles. The process of training the neural network is burdened with huge
calculation work and demand for big training datasets. However the use of already trained neural
model is fast. Therefore, the main obstacle in this field is the requirement to use huge training
datasets and training time.

Chemical Reactions Engineering Innovative Project

Uploaded by

Copyright:

Available Formats

Chemical Reactions Engineering Innovative Project

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Chemical Reactions Engineering Innovative Project

Uploaded by

Copyright:

Available Formats

Chemical Reactions Engineering

One of the central elements of the

The method presented in fig 3 reflects the

Data on the activation energy of chemical reactions

– iteratively performs predictions on randomly selected sets of training examples,

You might also like