Accelerating Materials Development via Automation, Machine Learning, and High-Performance Computing

Abstract

Successful materials innovations can transform society. However, materials research often involves long timelines and low success probabilities, dissuading investors who have expectations of shorter times from bench to business. A combination of emergent technologies could accelerate the pace of novel materials development by 10x or more, aligning the timelines of stakeholders (investors and researchers), markets, and the environment, while increasing return-on-investment. First, tool automation enables rapid experimental testing of candidate materials. Second, high-throughput computing (HPC) concentrates experimental bandwidth on promising compounds by predicting and inferring bulk, interface, and defect-related properties. Third, machine learning connects the former two, where experimental outputs automatically refine theory and help define next experiments. We describe state-of-the-art attempts to realize this vision and identify resource gaps. We posit that over the coming decade, this combination of tools will transform the way we perform materials research. There are considerable first-mover advantages at stake, especially for grand challenges in energy and related fields, including computing, healthcare, urbanization, water, food, and the environment. The development of novel materials has long been stymied by a mismatch of time constants (Figure 1). Materials development typically occurs over a 15-25-year time horizon, sometimes requiring synthesis and characterization of millions of samples. However, corporate and government funders desire tangible results within the residency time of their leadership, typically 2-5 years. The residency time for postdocs and students in a research laboratory is usually 2-5 years; when a project outlasts the residency of a single individual, seamless continuity of motivation and intellectual property is often the exception, not the rule. Market drivers of novel materials development, informed by business competition and environmental considerations, often demand solutions within a shorter time horizon. This mismatch in time constants results in a historically poor return-on-investment of energy-materials (cleantech) research relative to comparable investments in medical or software development. 1

A Critical Review of Machine Learning of Energy Materials

“…On the experimental side, high‐throughput measurements of materials optical properties are providing a large quantity of data that may be used for ML purposes . We are likely to see more efforts on experimental automation and high‐throughput works in the near future …”

Engineering 3D perovskites for photon interconversion applications

“…However, excitons rapidly dissociate in these materials, therefore this pathway is not expected to be efficient. Further investigation into Ruddlesden-Popper type perovskite-inspired materials [84][85][86] and less toxic double perovskites, combined with machine learning approaches [87][88][89][90] to find suitable materials has the potential to break the field wide open. In sensitized UC, not only the perovskite can be tuned, rather, we can also tune the energetics of the upconverting species.…”

Embedding physics domain knowledge into a Bayesian network enables layer-by-layer process innovation for photovoltaics

“…This corresponds to the progression from "Process Conditions" to "Materials Descriptors" in Figure 1. Device fabrication of solar cells is expensive, thus it is essential to explore the process variable space efficiently 27 . From a machine-learning point of view, we leverage the existing knowledge from literature and embed such domain knowledge as prior parameterization to constrain the parameter space e.g.…”