Scribd
Upload
10 views5 pages

Empirical Models

The document discusses the development of empirical dynamic models for chemical processes when theoretical models are impractical due to complex physics and difficult-to-measure parameters. It outlines a process for creating an empirical model through experimentation, model selection, and parameter identification, emphasizing the importance of model parsimony and the use of optimization algorithms to minimize loss functions. The Nelder-Mead simplex method is highlighted as a heuristic approach for optimizing model parameters.

Uploaded by

XheikhKaleem
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
10 views5 pages

Empirical Models

The document discusses the development of empirical dynamic models for chemical processes when theoretical models are impractical due to complex physics and difficult-to-measure parameters. It outlines a process for creating an empirical model through experimentation, model selection, and parameter identification, emphasizing the importance of model parsimony and the use of optimization algorithms to minimize loss functions. The Nelder-Mead simplex method is highlighted as a heuristic approach for optimizing model parameters.

Uploaded by

XheikhKaleem
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

CHE 361 developing empirical dynamic models

developing an accurate theoretical (i.e., physics-based) model of a chemical process based on phys-
ical principles could be impractical:

• many processes involve highly complicated physics

• often, there are many physical parameters (chemical & physical properties, reaction rate con-
stants, etc.) that are difficult/impractical to measure

alternative: develop an empirical (i.e., data-driven) model from process response data

process
empirical modeling is a generic term for activities that
u(t) y (t) create models by observation and experiment.
process
Wikipedia

eg. heterogeneous catalysis of gas-phase reaction A → B

∗
goal: find transfer function G(s) = CB (s)/F ∗ (s)
complications:

∴ it would be difficult and time-consuming to develop a physical model of this process.


→ let us instead develop an empirical model of the process.
1. conduct an experiment to interrogate the process about its dynamics

2. based on time series data characterizing the process response, select a suitable model

3. fit model parameters to the process response data


1. experiment. eg., observe the step response.

input output
f (t) [mol/s] cB (t) [mol/L]

t [s] t [s]

time series data characterizing the response of the process:

2. model selection.

propose the empirical process model:

CB∗ (s)
G(s) = =
F ∗ (s)

3. parameter identification.

we wish to determine the parameters:

→ tune the parameters until the model


cB (t; K, τ ) =
fits/matches the data

{(ti , cB,i )}ni=1 .


→ use G(s) (with the identified K , τ ) as the (empirical) process model.
...use G(s) for what?

process/system identification:

2
note on model selection

with four parameters I can fit an elephant, and with


five I can make it wiggle its trunk.

John von Neumann

Occam’s razor: in explaining something, make no more assumptions than necessary (invoked to de-
fend model parsimony and favor models with fewer parameters).

eg. consider the step response curve of a process:

y (t) model 1:

K
G(s) =
τs + 1

model 2:

t
K
G(s) =
a4 s4 + a3 s3 + a2 s 2 + a1 s + 1

ambitious Beavers: look up “Akaike information criterion” to select a model that appropriately bal-
ances model complexity (one metric = the number of parameters) and goodness of fit to the data.

defining model “fit” to data

build a cost/loss function that describes the “distance” between (i) the model with a given set of
parameters and (ii) the data. the sum of squares error is commonly used as a cost/loss function.

let y (t; β) be the dynamic model for the process response.

process response data: {(t1 , y1 ), (t2 , y2 ), ..., (tn , yn )}

the hope is:

[poorness of] “fit” quantified by the loss function:

ℓ = ℓ(β) :=

3
geometric picture:

y (t)

the loss ℓ:

minimizing the loss, ie., tuning the model parameters β to minimize the loss ℓ(β)

βopt =

how would such an optimization algorithm work?

• entire courses/careers on optimization. e.g. ECE 599

• in practice: use software such as Optim.jl in Julia

eg., the gradient-free Nelder-Mead simplex method: a heuristic algorithm to search for a minimum
of a function ℓ(β) by systematically transforming a simplex in parameter space. the idea is that the
simplex will eventually find, surround, and shrink in on the minimizer β opt .

a d -dimensional simplex can encapsulate the minimum β opt ∈ Rd .

simplices in 0, 1, 2, and 3 dimensions.

for first order systems β := [τ, K]

4
cartoon of the loss function ℓ(β) = ℓ(τ, K)

in the Nelder-Mead algorithm (the “dance of the simplices”), at each iteration, we choose a simplex
operation to perform:

reflection expansion contraction


β2 β2 β2

β1 β1 β1

a simplex operation is chosen, with the aim of enclosing the minimum, by sequentially:

• computing the value of the loss at each vertex of the triangle

• choosing the “best” side

• reflecting, expanding, or contracting the triangle, based on the loss at the newly proposed
vertex

You might also like