The Yaron Research Group

Machine Learning for Chemistry

A Layer for Deep Learning of Chemical Hamiltonians

Ab initio quantum chemistry predicts the properties of molecules by solving the Schroedinger equation for the motion of electrons in the molecules. This leads to accurate predictions but at high computational cost. One way to lower the cost is to generate a large set of quantum chemical data and then train a machine linear (ML) model that can faithfully reproduce that data. Once trained, the ML model may be used to make predictions at a cost that is orders of magnitude lower than ab initio quantum chemistry.

Diagram illustrating relations and equations for machine learning, quantum chemistry, and Density-Functional-Tight-Binding theory

The neural networks of deep learning are a particularly powerful form of ML model. Current neural networks use quantum chemistry only as a source of data. We are developing deep learning models that use quantum chemistry as an integral part of the prediction process. In particular, we have implemented self-consistent-charge Density-Functional-Tight-Binding (DFTB) theory as a layer for use in deep learning models. The DFTB layer takes, as input, Hamiltonian matrix elements generated from earlier layers and produces, as output, electronic properties from self-consistent field solutions of the corresponding DFTB Hamiltonian. Backpropagation enables efficient training of the model to target electronic properties. Once trained, the resulting DFTB Hamiltonian can be used to generate molecular orbitals, electron densities, energies, and other properties that are internally consistent because they all stem from the same model Hamiltonian.

Li, H., Collins, C., Tanha, M., Gordon, G. J., and Yaron, D. J. “A Density Functional Tight Binding Layer for Deep Learning of Chemical Hamiltonians” Journal of chemical theory and computation (2018)

Using Reinforcement Learning to Drive Chemical Reactions to Desired Outcomes

graph of Fraction of polymer chains (y-axis) and Reduced chain length (x-axis)

Reinforcement Learning (RL) is the subdiscipline of machine learning that enables computers to play the Go board game and to drive cars. RL is useful when the computer must make decisions, such as moving a stone in Go or stepping on a car accelerator, but does not get feedback on those decisions for quite some time. Once the results are in, such as winning the GO game or staying on the road, RL reinforces good decisions and penalizes poor decisions. Over time, the computer gradually learns how to make decisions that lead to desired outcomes.

We are developing ways to use RL to guide chemical processes to desired outcomes. As a first platform, we are using RL to control the shape of polymer molecular weight distributions (MWDs) in atom transfer radical polymerization (ATRP). The computer is given a set of actions, such as adding catalysts and other reagents, and the goal of achieving a specific distribution of polymer lengths (an MWD) once the reaction is complete. At present, the RL algorithm does its experiments on a reaction simulator, so that it can perform thousands of simulated reactions in a few hours. After some trial and error, the RL algorithm learns to make decisions that lead to a variety of target MWDs. We are now working on ways to transfer an RL controller, trained initially on a simulated reaction, to the physical laboratory.

Li, H., Collins, C. R., Ribelli, T. G., Matyjaszewski, K., Gordon, G. J., Kowalewski, T., and Yaron, D. J. “Tuning the Molecular Weight Distribution from Atom Transfer Radical Polymerization Using Deep Reinforcement Learning” Molecular Systems Design & Engineering 3, no. 3 (2018): 496–508.

Molecular Representations for Machine Learning


A crucial part of machine learning for chemistry is finding ways to represent the molecule as input to the machine learning algorithm. We have developed a new representation, that of encoded bonds, that helps models trained on smaller molecules to make predictions for larger molecules. We have also benchmarked a large variety of molecular representations to investigate the advantages and disadvantages of various approaches.

Collins, C. R., Gordon, G. J., Lilienfeld, O. A. von, and Yaron, D. J. “Constant Size Descriptors for Accurate Machine Learning Models of Molecular Properties” The Journal of chemical physics 148, no. 24 (2018): 241718.