Machine learning (ML) has the potential to revolutionize chemical discovery by replacing first-principles calculations with approximate property evaluation at negligible computational cost. It is increasingly well-established that ML models can learn the pattern of data obtained by first-principles methods such as density functional theory (DFT) to enable discovery of molecules and materials within the domain of the trained model. Still, most discovery workflows will require validation of these ML leads with computationally-demanding simulation. However, for some of the most promising materials (i.e., inorganic complexes and materials), two key challenges remain: i) we may be outside the domain of applicability of DFT due to the complex electronic structure of the material, or ii) a molecular structure may not form a stable complex. Typically, these two failure modes of simulation are determined after a computationally-costly calculation has been carried out. This challenge is compounded by the fact that such complexes would not normally enter into a training set or inform an ML model. To overcome these challenges, we trained a support vector machine on both failed and successful simulations to estimate simulation “promise“ (e.g., complexes that are stable and without spin contamination) directly from topological, heuristic features prior to simulation. These tools are integrated into a materials discovery workflow to inform which candidates are most worthy of study at the approximate DFT level. A dynamical classifier that utilizes the information of short-time DFT calculation is also developed, which aids the decision of whether to abandon an in-progress simulation, further optimizing the allocation of computational resources towards calculations most likely to be fruitful. Our classifiers are anticipated to be useful in accelerating the design of novel inorganic materials such as redox couples, spin crossover complexes, and catalysts in high-throughput screening.

Accelerating inorganic discovery with meta-calculation filtering via a decision classifier

256th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.