A neural network with embedded stoichiometry and thermodynamics for learning kinetics from reactor data

Abstract

The digitalization of chemical research and industry is vastly increasing the available data for developing and parametrizing kinetic models. To exploit this data, machine learning approaches are needed that autonomously learn kinetic models from large amounts of reactor data. In this paper we develop such a tool. We present a neural network architecture that embeds thermodynamic and stoichiometric prior knowledge (STeNN) for the accurate, robust and data-efficient modelling of chemical kinetics. This network architecture is used in conjunction with neural ODEs to autonomously learn kinetic models directly from reactor data. Using the example of an adiabatic steam reformer, we demonstrate that our approach accurately recovers the true kinetics from reactor data, where conventional neural networks fail. It is further shown that the proposed framework can handle large datasets and learns kinetic models from up to 1000 reactor experiments in around ten minutes. Furthermore, due to the embedded physico-chemical knowledge, our model is robust to significant noise in the data, even in the low-data regime. We anticipate that our approach, in combination with emerging big data frameworks, will greatly increase the availability of accurate kinetic models, providing a boost to model-based reactor design and control.

Goal‐Oriented Two‐Layered Kernel Models as Automated Surrogates for Surface Kinetics in Reactor Simulations

“…In contrast, surrogate models like splines or the (error-based modified) Shepard interpolation approach map precomputed solutions to accelerate reactor simulations [2,15,16,18,20,[27][28][29][30][31] or even spatial subsystems of the reactor [32,33] and breakthrough curves [34]. Lately, primarily machine learning techniques like random forests [35,36] or neural networks [2,3,37] have been used for accurate predictions of steady-state surface kinetics because they can overcome the so-called curse of dimensionality [38], i.e., the exponentially increasing difficulty to learn high-dimensional data. A promising alternative are kernel methods because their training is deterministic and data efficient even for highdimensional problems.…”

“…3.2. and breakthrough curves [34]. Lately, primarily machine learning techniques like random forests [35,36] or neural networks [2,3,37] have been used for accurate predictions of steady-state surface kinetics because they can overcome the so-called curse of dimensionality [38], i.e., the exponentially increasing difficulty to learn high-dimensional data. A promising alternative are kernel methods because their training is deterministic and data efficient even for highdimensional problems.…”

Robust Mechanism Discovery with Atom Conserving Chemical Reaction Neural Networks

“…In this work we focus on implementing the atom balance in neural networks. Previously, this has either been done explicitly through a soft constraint in the loss function [4,5], or a post-processing step [6], or indirectly by embedding the stoichiometric matrix into the model's structure [7][8][9][10]. However, in case of mechanism discovery and reduction, the stoichiometric matrix is generally unknown and subject to optimization.…”