MRS 2020 #2: Predicting materials synthesis
tags: [dft mrs machine_learning Being able to predict the properties of a hypothetical material is one thing, being able to predict if that material can ever be realised is quite another thing. I can recall hours playing around with swapping elements and playing what-if games in the DFT universe. You could discover all kinds of great new electronic structures, but show them to an experimental chemist and they would laugh you out of the room. This actually highlights two problems, one (probably the main one) is that simply being able to optimise a geometry is no guarantee that a structure is actually plausible, but the other is that experimental chemists tend to be very focussed on the space of compounds and molecules that they know, it can be hard to discover truly novel systems. So some kind of AI that can reliably and in an un-biased manner explore chemical space should be able to overcome this, right? The problem is how do measure synthesisability to train such a system?
Chris Bartel gave a nice talk on how many current ML approaches fail badly at predicting the basic requirement for synthesisability of a material, that is its stability with respect to decomposition. Almost all composition-based ML models produce deformation energies that are not much better than random guessing. Including some knowledge of structure in the ML models, e.g. graph networks, improves the situation considerably.
- Talk: Contrasting Error Cancellation in Ab Initio and Machine Learning Predictions of Thermodynamic Stability
- Paper: A critical examination of compound stability predictions from machine-learned formation energies
Evan Spotte-Smith presented some impressive work on using graph data structures called computational reaction networks. By casting various reactions as a connected graph problem you can then use (with some domain specific modifications) the many existing tools from computer science (like the PageRank algorithm that powers Google) for efficiently traversing and searching complex graphs.
- Talk: Automatic Generation of Computational Reaction Networks for Unbiased Exploration of Chemical Pathways
- Paper: A Chemically Consistent Graph Architecture for Massive Reaction Networks Applied to Solid-Electrolyte Interphase Formation
A slightly different problem is addressing how to modify the synthesis conditions to achieve a desired outcome. There were a couple of really impressive talks about using reinforcement learning and related techniques to optimise dynamic control parameters e.g. temperature, pressure, concentrations during a synthesis to achieve an optimal goal.
I found the work from Pankaj Rajak on reinforcement learning for optimising CVD of $MoS_2$ a really impressive piece of work. They developed LSTM models based on molecular dynamics simulations to model how a system outputs (percentage $MoS_2$ formed, defects etc) evolves based on the current states and the control parameters. The LSTM is then coupled with reinforcement learning “Agent” which makes suggestions on how to update the control parameters. There is then a reward depending on how good the state of the system is judged to be after the updated and this reward is used to modify the agent policy and based on the new system state and the new policy the agent makes a new suggestion. It’s a shame that so far this was only demonstrated on reproducing molecular dynamics simulations, but an impressive proof of principle nonetheless.