deng-mit / crnn

Chemical Reaction Neural Network

CRNN (Chemical Reaction Neural Network)

CRNN is an interpretable neural network architecture for autonomously inference chemical reaction pathways in various chemical systems. It is designed based on the following two fundamental physics laws: the Law of Mass Action and Arrhenius Law. It is also possible to incorporate other physics laws to adapt CRNN to a specific domain.

You can find the common questions regarding CRNN in the FAQs.

Structure of this repo

This repo provides the case studies presented in the original CRNN paper as well as ongoing preliminary results on other systems. Currently, we are actively working on the following systems:

• Biomass pyrolysis kinetics
• Cell signaling pathways for quantitative modeling drug effects
• Gene regulatory network (preliminary results)
• Oscillations in yeast glycolysis (preliminary results)
• Systems with strong stiffness (Robertson's problem)
• Gas-phase combustion kinetics (HyChem)

Inside each folder, such as case 1/2/3, you will find at least two Julia codes. One for training and the other one for weight pruning. Those two files are identical, except that the weight pruning includes a function to prune the CRNN weights to further encourage sparsity.

Get Started

Have a look at the code for case 2. The script consists of the following parts:

• Hyper-parameter settings
• Generate synthesized data
• Define the neural ODE problem
• Train CRNN using ADAM

We strongly recommend using Julia 1.6 with the CRNN code included in this repository. Newer versions may lead to indexing issues and convergence to incorrect mechanisms.

Cite

Ji, Weiqi, and Deng, Sili. "Autonomous Discovery of Unknown Reaction Pathways from Data by Chemical Reaction Neural Network." The Journal of Physical Chemistry A, (2021), 125, 4, 1082–1092, acs/arXiv

crnn's Issues

A stiff problem Robertson’s chemical reaction model

Plan to run a stiff problem to synthesis the difficulty of combustion problems.

Robertson's chemical reaction model is typically used to test algorithms for stiff problems.

http://www.dam.brown.edu/people/alcyew/handouts/numODE5.pdf

Guidelines for bounds on clamping weights , scaling derivatives and slope

Opening issue for bit detailed documentation on

• slope,
• scaling dydt,
• bounds on clamping weights, scaling weights
• scaling loss by the standard deviation
• sensealg for different cases (stiff and non-stiff)
• when to bother about g_norm or gradient in the training process(like we don't bother much in simple case 1)
• tanh vs gelu ? Comments on different activations in different cases

Query in robertsons code in predict_neuralode function

Hi @jiweiqi ,In robertson's code.. shouldn't one solve for _prob and not prob .. copying here code for the reference from
https://github.com/DENG-MIT/CRNN/blob/main/robertson/rober_crnn.jl

function predict_neuralode(u0, p; sample = datasize)
    global w_in, w_b, w_out = p2vec(p)
    _prob = remake(prob, tspan=[0, tsteps[sample]])                               # this line
    sol = solve(prob, alg, u0=u0, p=p, saveat=tsteps[1:sample],             # this line
                sensalg=sense, verbose=false, maxiters=maxiters)

``

creating a new issue for possibility of optimising CRNN using Optim library optimisers

https://discourse.julialang.org/t/ann-fluxopttools/25888
This link talks about using other Optim library optimisers along with Flux.

Biofuel from Palm Oil or Other Oils

Social impact of Palm Oil

Palm oil is a small ingredient in the U.S. diet, but more than half of all packaged products Americans consume contain palm oil—it’s found in lipstick, soaps, detergents, and even ice cream.
https://rspo.org/about
https://www.worldwildlife.org/industries/palm-oil

Ideas on training CRNN from experimental data

It would be nice to study a case with experimental data. The chemical kinetic model can be very simple, although. It would be useful when you show something to others.

Maybe Ritaank and Matthew can help with it. I will also make up some candidates on this issue.