Lin Du.

The agent proposes bounded design parameters. A Python controller checks the schema, geometry and feasibility before a fixed Java/COMSOL implementation constructs and solves the model.

Each attempt retains its proposal, validation outcome, solver log, raw exports and derived metrics. Isolated solver processes and explicit execution states make failures inspectable and iterations reproducible.

Outcome: The selected design increased reservoir area from 46.18 to 49.38 mm² while reducing surface t90 from 32.01 to 29.34 minutes. This remains a simulation-based proof of principle; physical validation and autonomous laboratory execution are future work.

Extract without peak annotations: gradient-pattern clustering locates peak regions. Polynomial baseline subtraction isolates the peak current, and GMM decomposition estimates peak location and width before amplitude scaling.

Build a shared representation: each peak becomes an amplitude, position and width triplet, independent of raw sample count. Concentration-dependent amplitude modelling accounts for pairwise activation or inhibition.

Generate with experimental constraints: modelled concentration trends are combined with residual variability and observed parameter ranges to generate plausible signal families. Full dataset generation and decomposition of overlapping peaks were evaluated in complementary experimental settings.

Measurement → representation: peak detection, baseline subtraction and Gaussian/GMM feature extraction convert cyclic voltammograms into structured model inputs.

Representation → prediction: Optimized IdentQuantNet combines identification with drug-specific, multi-branch quantification. Interaction-aware modelling and tailored asymmetric losses account for interference and unequal error costs.

Validation: simulated and independently measured datasets, with ETO, MTX, IFO, CP and 5-FU across the studied single-agent and interacting-drug cases. Read the signal-simulation paper

Define a plausible domain: population PK models, patient covariates and dose history generate prior concentration profiles. Monte Carlo population sampling and percentile bounds define the evaluation domain independently of a particular measurement.

Separate decision from consequence: observed concentration and recorded sampling time drive Bayesian estimation and regimen selection. The selected regimen is then evaluated under the corresponding true PK state, and the resulting exposure is assigned to one of five therapeutic risk categories.

Repeat across uncertainty scenarios: concentration and timing discrepancies are evaluated systematically. Time-conditioned slices visualise the resulting robustness space. Drug-specific case studies show that sampling phase, uncertainty source and therapeutic target can change robustness; informativeness and robustness are distinct objectives.