Reframing Research as Optimization

At the core of our autonomous discovery initiative is the realization that many complex materials research challenges can be fundamentally reframed as optimization problems. By carefully constructing and deploying reward-based systems, our group is building a broad, versatile suite of automated optimization workflows capable of accelerating materials discovery.

A critical aspect of this paradigm shift is our transition from traditional myopic (short-sighted, grid-based) measurements to non-myopic active learning. Rather than passively scanning an entire sample, our algorithms intelligently predict and target the most valuable regions to explore next. This reward-driven approach maximizes information gain, minimizes experimental time, and allows us to actively steer the instrument toward specific physical states.

Our Core Optimization Workflows:

  • Instrument Optimization: Automated tuning, alignment, and parameter optimization for complex instruments, including both STEM and SPM.
  • Combinatorial Libraries: High-throughput, non-myopic exploration and mechanical/structural mapping of structured combinatorial material libraries.
  • Random Libraries: Rapid property screening and phase discovery across complex, unstructured random material spaces.

Reward-driven STEM

Contributors: Kamyar Barakati, Utkarsh Pratiush, Austin Houston, Gerd Duscher

A demonstration of reward-driven optimization utilized for real-time instrument tuning and alignment.