Surrogate model in IDAES

[abhi24x7]

I have been familiarising myself with the framework now, and documentation is very elaborate and descriptive. Thank you IDAES community.

Using the IDAES framework, I now have modelled a packed bed reactor using customized non-Arrhenius rate equations, now I aim to model the same reactor using a ANN based surrogate model, embed it into the flowsheet inplace of the PFR block and run optimization over the flowsheet.
For getting started with ANN based surrogate models with IDAES, I have referred
https://idaes.github.io/examples-pse/latest/Examples/SurrMod/FlowsheetOptimization/Keras_flowsheet_optimization_doc.html

However this is just illustrating the ANN based model for a single unit, i.e. reformer. It doesn't have any other unit operations ( and related physical property package) in the flowsheet. Please correct me if I am wrong.

I was wondering if embedding a ANN based model for the Packed bed reactor is possible in IDAES (assuming it has to be linked with the physical property package in someway), any support would be highly appreciated. Please let me know if further clarity is needed.

Thank you very much in advance!

[andrewlee94]

@abhi24x7 It is definitively possible to embed an ANN based model into a flowsheet, but there are a few ways to do it. Based on your comment above, it looks like you are thinking about creating an ANN for the entire PFR with reactions (which is a good idea if you have complex reactions).

Firstly, your ANN based model does not have to be linked to any specific property package. IDAES supports having different units in the process use different property packages, or even no property package at all if it makes sense. The one thing your surrogated-based model will have to do is declare Ports which will be used to connect to the other units in the flowsheet.

Next, when it comes to connecting the Ports to other units in the flowsheet, the Ports need to have the same contents; basically they have to have the same set of variables and indices and consistent references states. Using the same property package in each unit model normally ensures this, but it is also possible to "translate" between different property packages (or Port definitions).

So, in your case you have two options.

1. Define the Ports in your ANN based model to be consistent with the property package you are using up- and down-stream. This is probably easy in your case, as you will be building the ANN from a PFR using the same property package.
2. Alternatively, IDAES also has a TranslatorBlock which allows you map the Ports in one property package to those in another. Basically, this is just an empty unit model with an inlet Port defined using one property package and the outlet using the other, and it is then up to you to write a set of constraints which do the mapping (hint: if you do this, always equate temperature and not enthalpy as temperature is absolute whereas enthalpy is based on a reference state).

[abhi24x7]

Dear @andrewlee94,
Thank you very much for your reply and assurance that ANN based PFR is possible, could you please tell me if there is any resource/example (apart from the AutoThermalReformer surrogate model), that shows how to embed a ANN based PFR inplace of PFR block in a flowsheet?

Secondly, I am planning to make the ANN using Pytorch, is it possible to use a pytorch model in IDAES or do I have to make a keras model like the one describe in the AutoThermalReformer example?

[andrewlee94]

@abhi24x7 Here is an example of using the IDAES SkeletonUnitModel to implement a surrogate model into a flowsheet: https://idaes-examples.readthedocs.io/en/latest/docs/unit_models/operations/skeleton_unit_doc.html

At the moment, I think we only support Keras (although I am not sure as I do not help maintain those tools).

[abhi24x7]

Hello,
Thank you for all your previous replies!

With regards, to my ANN based PFR model. I am slightly stuck in figuring out a method to embed it into IDAES, for better explanation from your end, I would like to explain you about the ANN model:

Inputs : length of reactor (z), inlet Temp, inlet Pressure, inlet feed flowrate of CH4, inlet feed flowrate of O2, inlet feed flowrate of CO2, inlet feed flowrate of of H2O.

Outputs : outlet Temp, outlet Pressure, inlet feed flowrate of CH4, inlet feed flowrate of O2, inlet feed flowrate of CO2, inlet feed flowrate of H2O

Since the variable "z" is not a state variable that can be transferred from one unit to another, how do I deal with this issue?

While going through the IDAES documentation I came to know that, ANN type surrogate models can be embedded using OMLT (as shown in the ATR example). But there is only one unit and no other unit operations. The reason I am stating this is because, I am planning to include other units like compressors that use a property package with state variables as: FTPx
How can I make the inputs and outputs of the ANN model consistent with the property package that is going to be used in other unit blocks?

Thank you and looking for your reply!
Thank

[andrewlee94]

@abhi24x7 When you embed your ANN into IDAES, the want to map all your ANN inputs and outputs to Pyomo Vars. You can then decide how to expose all of these to the user; the feed and outlet stream will be mapped to Ports so that they can be connected up and down stream, but you can have other variables (such as z) that are not included in the Ports and are treated as Vars the user is expected to fix themselves (e.g. volume in the CSTR unit model: https://idaes-pse.readthedocs.io/en/stable/reference_guides/model_libraries/generic/unit_models/cstr.html?highlight=CSTR).

[abhi24x7]

@andrewlee94,
Thank you very much for your prompt reply and suggestion.

First and foremost, I apologize if I have repeated certain things from the previous replies, its just that I want to understand the development of the model in a better manner and I have been facing slight difficulty in doing so.

Having said that, I plan to use OMLT framework to include my trained ANN into Pyomo formulation, and the ANN is trained to accommodate the inputs and outputs in the following order:

Inputs : length of reactor (z), inlet Temp, inlet Pressure, inlet feed flowrate of CH4, inlet feed flowrate of O2, inlet feed flowrate of CO2, inlet feed flowrate of of H2O.

Outputs : outlet Temp, outlet Pressure, outlet feed flowrate of CH4, outlet feed flowrate of O2, outlet feed flowrate of CO2, outlet feed flowrate of H2O.

So based on your reply, when you say , "map all your ANN inputs and outputs to Pyomo Vars", does it mean that I make a different python script ANN_PFR.py (different from where I have defined the rest of the flowsheet) where in load the trained ANN model as an OMLT formulation?
Or do I load it in the Jupyter notebook where the rest of the flowsheet is established, like how it is shown in this

Secondly, I had a look at the CSTR unit model source code and source codes of other unit models as well, I see that all of them have a CONFIG = UnitModelBlockData.CONFIG() which stores Balance equations (material, energy & momentum) which are called in the build() method on the ControlVolume0D which makes sense.

However in my case do I need to implement the CONFIG = UnitModelBlockData.CONFIG() as well? Because I don't see why I should implement the balance equations, please correct me if I am wrong.

Another thing is that in the rest of the flowsheet, the state variables that are being passed are FTPx, I cannot understand how can I directly link the ANN model to the other units in the flowsheet as these are not the variables "directly" being computed by the ANN, so I wonder how do I pass the FTPx state variables.
One indirect way I thought about this is to write a custom function that takes in the FTPx values from the previous unit, converts (as well as scales it) to the type of input that the ANN requires and once predicted, another custom python function would unscale the predicted values and compute the output FTPx as required by the rest of the flowsheet.
I feel that there might be a more smarter way to think about this, because based on the current description, I might run into issues/errors beyond my understanding.

In the CSTR unit model source code I see that the ports are added by self.add_inlet_port() and self.add_outlet_port() and I presume that these ports read the inlet and outlet state variables as defined by the CONFIG. I also wanted to know how can I pass the variables computed by the ANN that the ports can easily transfer to the other units of the flowsheet.

Thank you for reading through my issue, and apologies for the long text I just tried to describe it in the best possible manner I can.

Looking forward to your reply.

[andrewlee94]

1. From that error, it appears that you have a mismatch between the indexing set for .mole_frac_comp_in and the initial values you are providing it. Notably, it looks like you have provided an initial value for carbon_monoxide, but Pyomo is telling you that is not in the indexing set. Start by checking the indexing set to make sure carbon_monoxide is there and is spelt correctly.
2. That line of code is creating a new Pyomo Var, indexed by the flowsheet time domain (m.fs.time), the phase list from the associated property package (m.fs.pervap.phase_list), and component list from the property package (m.fs.pervap.comp_list).
3. Without seeing your code, it is hard for me to say if you've done everything correctly or not, and even then it might be hard to interpret everything your are doing. The place to start is by using the Pyomo display() method to see if those variables have been added to your
   model, and if they have the correct indexing sets and values.
4. You are not declaring the indexing sets again when you create the variable, you are telling Pyomo that these sets are to be used to index this variable. Note that you are passing the existing instances of these sets to the Var() class, so they already exist.

[abhi24x7]

Dear @andrewlee94,

Thank you for your quick reply! I understand all your reasoning, with respect to point 3, is there any section of code you want to specifically look at? I have also attached the code block in the previous reply as well named "Surrogate_model_v1.txt.txt"

[andrewlee94]

Ah - I see the problem. The issue is that when you initialize the Vars, the dict you need to provide needs to match the full indexing set for the Var. I.e., carbon_monoxide is not a valid key as m.fs.sm.mole_frac_comp_in is indexed by time, phase and component. What you need is:

m.fs.sm.mole_frac_comp_in = Var(
    m.fs.time,
    m.fs.sm.phase_list,
    m.fs.sm.comp_list,
    initialize = {
       (0, "Vap", 'carbon_monoxide'): 1e-2,
       (0, "Vap", 'carbon_dioxide'): 1e-2,
       (0, "Vap", 'hydrogen'): 1e-2,
       (0, "Vap",  'methane'): 1e-2,
       (0, "Vap", 'methanol'): 1e-2,
       (0, "Vap",  'water'): 1e-2,
       (0, "Vap", 'nitrogen'): 1e-2},
    units = pyunits.dimensionless)

One other thing to note on naming and conventions in IDAES: the name mole_frac_comp_in indicates that this is mole fractions indexed by time and component list, however you also included the phase list in the indexing sets. I am not sure which you actually want (this will depend on the up- and down-stream property packages, but if you need to include the phase as an index then the standard name would be mole_frac_phase_comp.

[abhi24x7]

Dear @andrewlee94,

Thank you very much for your previous reply.
I was also looking at other different ways to embed the surrogate model in IDAES, as I was not gaining much success with SkeletonUnitModel. I came across the source code of the scientific article titled Process Flowsheet Optimization with Surrogate and Implicit Formulations of a Gibbs Reactor

Link to article: https://arxiv.org/html/2310.09307v2
Link to source code: (https://github.com/Robbybp/surrogate-vs-implicit/tree/main)

In this work, they also use a Neural network trained in Keras and they embed it into IDAES. This something very similar to what I aim to do in my case, with the difference being, here they use a Gibbs Reactor as the underlying reactor model and I use the PFR. Even they are using the state variables FTPx.

When I look at the way they have implement the "embedding" of the ANN based model (link to the code), I found something interesting, firstly, they haven't used any SkeletonUnitModel,

• They have declared a SurrogateBlock,
• declared the inputs and outputs variables (as a python list)
• And finally they have loaded the trained ANN model and have build the formulation.

Before they have done the above said steps they have done this:

From what I understood when I looked up, what this could do is that, the above lines "deactivates" all the constraints related to the original Gibbs reactor. Is this right? Additionally, what I also assume that once the constraints are deactivated, it essentially "by-passes" the original Gibbs reactor calculation routine, which is critical to shift it to the predictions by the Surrogate model.

Secondly, I noticed that the outputs of the ANN are directly referenced from the variables of the original IDAES Gibbs Reactor, like as shown:

What is the need to do the same?

Now, coming to my case
(Simulating just the ANN using OMLT, no integration with IDAES)
I am using an onnx file of a neural network (about which I elaborately describe in my previous replies). I then use a OmltBlock to make it to a formulation. And I then connect the inputs that are already in the form which the neural network needs, I solve it using a solver and it readily predicts the outputs!

(Integration of ANN with IDAES)
To keep it simple, I just have a Feed, PFR and Product block, I have declared all the Degrees of Freedom for all the unit blocks, connected the arcs and have Initialized it. (Things are good till here!)

Now to implement the ANN in the IDAES, I tried to implement a similar approach to what is done in the Gibbs ATR example in my case, where I deactivate all the constraints related to my PFR and then declare the respective output variables, as shown

After this, I declare the scaling factors (to scale the inputs for the ANN and to unscale outputs from the ANN) as Pyomo Param. I then connect the scaled variables to the inputs of the ANN OmltBlock
Kindly note that: my ANN takes length of reactor (z), T, P and molar flowrates of the component (I calculate this by multiplying flowrate by mole fraction and pass it to NN)

Similarly, I connect the scaled outputs of the Neural Network, to the scaling factors to unscale it and pass it to the state vars of R101

After this when I attempt to solve the flowsheet, I get the following error:

When I looked into the model using m.fs.nn.display(), I see that

The inputs have been scaled properly and fed as the input to the ANN, however the outputs are simply 0 (which somewhat related to the error the solver is throwing.)

This is the problem I am facing, I tried to elaborate it as much as possible and I have been trying to debug but couldn't reach any further. Could you please suggest me what can be done here?

Thank you and looking forward to your reply!!

[andrewlee94]

So, first of all if the error you are getting is a solver failure, and notably an infeasible solution. This means that the code built the model successfully, but that it was unable to solve which means the issue lies in the formulation of the model and not in the syntax of the code.

The first place to start with problems like these is the Diagnostics Toolbox:

Documentation: https://idaes-pse.readthedocs.io/en/stable/explanations/model_diagnostics/index.html
Tutorial/Example: https://idaes-examples.readthedocs.io/en/latest/docs/diagnostics/diagnostics_toolbox_doc.html

These tools will help you look at your model and try to work out where the problem lies. Looking at the error message you are getting, a few things jump out:

1. It failed in zero iterations.
2. It says that there is an infeasibility in the triangular part of the problem. This means you have a system of equations in the model that can be solved sequentially (e.g. A = B, B=C, etc.) which is infeasible - this is probably not related to the ANN itself but the other equations in the model.

You should start by first making sure the degrees of freedom in the model are 0 (non-zero degrees of freedom can mask underlying issues) and then check the bounds on all your variables. It is likely you are trying to set a value outside a bound and this is causing the infeasibility. The Diagnostics Toolbox can help you identify these,