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

Machine learning (ML) has the capacity to revolutionize materials discovery with accurate property estimation at negligible computational cost. Still, most discovery workflows require computationally-demanding simulation to generate data to feed in an ML model. However, two key challenges remain at the stage of data generation: i) materials may not form a stable complex and ii) calculations may fall outside the domain of applicability of the chosen method. Usually, these two failure modes can only be detected after calculations finished, leading to a massive waste of computational resources. To address this problem, we trained a classifier to estimate the success rate of a calculation directly from topological, heuristic features prior to simulation. Inspired by the data distribution in the latent space, we designed a model confidence metric specifically for classification tasks, lowering the risk of terminating jobs that are actually fruitful. A dynamical classifier that utilizes the information generated during simulation is also developed, which directs the on-the-fly decision of whether to abandon an in-progress calculation. Our classifiers are useful in dataset generation with first-principles calculations to accelerate the ML-driven design of novel inorganic materials.