Apr 23, 2024
11:15am - 11:30am
Room 320, Level 3, Summit
Matthew Carbone1
Brookhaven National Laboratory1
Optimal design of experiments is an outstanding challenge in the materials science community. Standard techniques such as Bayesian optimization can address this challenge as long as the goal of an experimental campaign is maximizing some observable. While useful in a wide variety of cases, it is insufficient to address more epistemic goals, such as thoroughly exploring a sample space when an explicit objective is lacking. In this work, we showcase a formulation of "scientific value," a scalar representation of local uniqueness. Our method scalarizes arbitrarily high-dimensional data, allowing for the traditional optimization toolbox to be applied to any problem. Intuitively, our method, which we call the Scientific Value Agent, explores regions of space in which observables change rapidly, whilst exploring other areas sufficiently, wasting minimal experimental budget. It is also robust to the cold start problem, and can be utilized with no prior knowledge and next-to-no data. We demonstrate this technique by exploring a variety of simulated and real-world examples, including phase boundaries, autonomously changing the temperature measurement of a ferroelectric material, and analyzing nanoparticle synthesis data. Our method can be seamlessly extended to studies with multiple outputs, or fidelities, and can seamlessly integrate into existing autonomous experimentation frameworks with minimal added effort to the user.