Abstract
Silver nanoparticles are widely applied in medicine, packaging, and environmental remediation because of their potent antimicrobial properties. Nevertheless, predictive modeling of their zone of inhibition (ZOI) remains a longstanding challenge due to limited experimental datasets and the lack of physical consistency in machine learning approaches. In this study, a physics-guided liquid state machine (PG-LSM) was developed, integrating key physics-informed features of nanoparticle formation with experimental ZOI data through a series of reservoirs to predict antimicrobial efficacy. Three pretrained reservoirs of particle shape, particle core size, and ultraviolet-visible peak are transferred as encoders into a terminal ZOI predictor. Across nine baseline models, PG-LSM achieved the strongest test performance with a coefficient of determination (R²) of 0.956, a root mean square error (RMSE) of 1.151, and a mean absolute error (MAE) of 0.486 with close agreement between validation and test scores, indicating reliable generalization under small-sample conditions. Ablation studies confirmed the additive value from each reservoir; removing any intermediate stage degraded accuracy and stability. SHAP analysis revealed that exposure dose concentration and duration dominated the antimicrobial activity of AgNPs, whereas microbial species, capping agent, and reducing agent had a lower impact. PG-LSM advances materials informatics for AgNP antimicrobial assessment through physics-guided embeddings integrated with reservoir computing. The framework delivers accurate, stable, and interpretable ZOI predictions and remains practical for limited datasets.