Abstract: Complex material systems play a growing role in addressing some of our current technological challenges. They include multi-principal element alloys and oxides (often also called high-entropy alloys oroxides), metastable materials such as metallic glasses and carbon nanotubes, solid-state energy storage devices, heterogeneous catalysts, and increasingly complex devices for power electronics and quantum computing. However, there are easily tens of millions and possibly even a billion compositions out of which a very small fraction are potential candidates. Many of the desired materials are metastable, and therefore, processing paths play a crucial role in synthesizing them. The composition-processing combinatorial search space is too vast for a dependence on serendipitous discoveries to provide urgently need novel materials and devices; and a brute-force survey of the combinatorial space is too slow and expensive. We need guidance to navigate the vast composition-processing space. Our physiochemical understanding of complex material systems is still too nascent to provide it.

Recent advances in machine-learning (ML) and other artificial intelligence (AI) tools, however, suggest that a data-driven approach that builds on insights from physiochemical theories and bridges gaps in insights by experimental observations can provide the needed navigation. Application of ML to large experimental observations can highlight hidden and complex trends that theories may miss and help direct us to the next high-impact experiment to perform, leading to accelerated discoveries and new physiochemical insights. Iterations of a ML-guided cycle, starting from predictions, to smart experiments, to new discoveries and insights, resulting in the next generation of predictions, is an approach that would not only significantly accelerate the pace of discoveries but will also change the way we do science.

The data-driven approach needs a large amount of data, which fortunately, the last 30 years of public investment in national scientific user facilities is primed to produce. The rate and complexity of measurements at these facilities has been increasing exponentially over the last two decades. However, for the -driven discoveries, we do not need raw measurements; we need the new scientific information (knowledge) contained in them. Unfortunately, the rate of new knowledge lags the rate of measurement because it is extracted by humans, who are struggling to keep up as the pace and complexity of measurements is accelerating. We need new data-tools to produce knowledge from measurements in real-time. We also need tools that utilize the newly produced knowledge to make experimentation smarter. Operation of the data-driven discovery cycle also needs a supporting cyber-infrastructure that provides on-demand computation and seamlessly movement data where and when needed.

Here, I will use the search for new multi-principle-element alloys, and specifically wear-resistant metallic glasses, to illustrate the data-driven discovery paradigm. I will also use it to highlight the analytical and cyber-infrastructural challenges that must be addressed for this paradigm to produce the urgently needed new materials and technologies.