Agents for sequential learning using multiple-fidelity data.

Area of Science:

• Computational Materials Science
• Machine Learning for Materials Discovery
• Chemical Informatics

Background:

• Sequential learning optimizes data acquisition in materials discovery for exploration or exploitation.
• Real-world campaigns face costly data acquisition, necessitating strategies using varied data fidelities, like computational (e.g., DFT) and experimental data.

Purpose of the Study:

• Introduce novel agents capable of operating with multiple data fidelities for materials discovery.
• Benchmark the performance of these multi-fidelity agents in an emulated campaign targeting materials with specific band gap values.

Main Methods:

• Developed agents that integrate data from different fidelity sources: Density Functional Theory (DFT) calculations (low fidelity) and experimental results (high fidelity).
• Evaluated agent performance on an emulated materials discovery campaign focused on achieving target band gap values.

Main Results:

• Demonstrated performance gains by agents incorporating multi-fidelity data.
• Showcased benefits in two contexts: using low-fidelity data as a prior knowledge base and acquiring low-fidelity data concurrently with experimental data.

Conclusions:

• Multi-fidelity sequential learning agents enhance the efficiency and discovery rate in materials science campaigns.
• Provides a valuable tool for optimizing data acquisition strategies and hyperparameters in practical, multi-source active learning scenarios.