Chemistry-informed Bayesian control and policy distillation for autonomous pH adjustment in diverse chemical systems

Abstract

Adjusting pH of an arbitrary solution to a specified target remains difficult to automate when buffering behavior is unknown, nonlinear, and composition-dependent. We formulate pH adjustment as a sequential inference-and-control problem and develop a chemistry-informed Particle Filtering (PF) Bayesian controller that updates an explicit dose–pH response model. Across five independently generated benchmark sets of 3000 simulation tasks, the PF controller achieved 95.4 ± 0.6% success, requiring 4.84 ± 0.09 additions among successful tasks. PF state-action trajectories were subsequently distilled into a lightweight neural policy, which achieved 89.2 ± 0.5% success. In a separate matched 100-task timing cohort, the median observation-to-action latency was 40.1 ms per recorded PF decision cycle and approximately 0.15 ms per neural-policy call. Proximal Policy Optimization (PPO), a reinforcement-learning method, was then evaluated as an optional refinement of the distilled policy and achieved 93.9 ± 0.6% success with 5.2 ± 0.1 additions among successful tasks. The PF and PPO controllers were then implemented on an automated liquid-handling and pH-sensing platform and applied to mixed-acid solutions, laboratory wastewater, pH-controlled milk acid precipitation, and Cu(ii)–sulfosalicylic acid complexation. Together, these results establish a workflow connecting inspectable online response-model inference with policy distillation and optional data-driven refinement for autonomous pH control.