Accelerating sustainable glass discovery: integrating molecular dynamics, machine learning, and robotic synthesis

Abstract Discovering sustainable glass compositions demands navigating vast chemical spaces—a challenge that conventional experimentation cannot meet efficiently. Here we introduce the Simulation-Calibrated Active Learning Estimator (SCALE), a closed-loop framework uniting high-throughput molecular dynamics (MD), machine learning (ML), and robotic synthesis to bridge the gap between simulation and experiment. Over 42,000 melt–quench MD simulations with compact cells ( ≈ 200–500 atoms) train ML models that predict density and elastic moduli with cross-validated R 2 values up to 0.98. Comparison with 55 robotically synthesized sodium alumino-borosilicate glasses reveals systematic density overestimation of up to 5%. Rather than naively augmenting training data, SCALE iteratively learns a composition-dependent calibration from minimal targeted measurements. The protocol substantially reduces density errors over three experimental iterations of six measurements each and demonstrates the potential for glass optimization with a small number of strategically chosen experiments.