Although our work spans multiple application domains—including pharmacokinetics, epidemiology, and health economics, the underlying methodology is domain-agnostic.
Across all applications, we work with systems governed by dynamical structure, partial observability, and data limitations. Our approach emphasizes:
mechanistic representations of system dynamics,
identifiability-aware inference under structural constraints,
characterization of admissible model and parameter sets, and
propagation of uncertainty to predictions and decisions.
This framework applies equally to molecular-scale transport models, population-level disease dynamics, and economic decision models. Differences between domains enter through model structure and data, not through changes in the underlying inferential principles.
Many scientific and regulatory problems are governed by mechanistic structure: differential equations, conservation laws, and coupling across scales. In such settings, data alone cannot uniquely identify a single “best” model.
Our approach treats mechanistic structure as a form of inductive bias. Rather than relying on point estimates, we characterize admissible sets of models and parameters that are jointly consistent with data, governing equations, and validation criteria.
We use Bayesian and optimization-based inference to preserve parameter correlations, quantify non-identifiability, and propagate uncertainty to predictions. This allows extrapolation beyond observed data while maintaining transparency about what is, and is not, learnable from available information.
The Model Master File (MMF) framework formalizes validated mechanistic models as reusable regulatory assets. An MMF represents not a single calibrated model, but a structured family of admissible models supported by data and mechanistic constraints.
Our work on MMF development has been informed by invited participation in FDA-led workshops and international presentations on MMF applications for dermatological products. We view MMFs as a general data-science construct for reusable, auditable, and extensible dynamical models.