Ensemble-embedded physics-informed neural networks
Researchers: Dr Kanthu Joseph Mhango; Megan J. Lewis; Joe M. Roberts; W. Edwin Harris
This project develops physics-informed neural networks (PINNs) that embed established empirical and ensemble approaches inside an end-to-end trainable neural architecture. The objective is to retain structured scientific knowledge and diversity of model perspective while learning nonlinear, context-dependent behaviour from data.
Scientific challenge
Scientific disciplines often contain several useful empirical models rather than one universally adequate formulation. Conventional ensembles can combine those perspectives robustly, but they commonly remain separate from the internal representation learned by a neural network. Unrestricted neural models offer flexibility yet may ignore known structure, recover scientifically implausible relationships or make it difficult to determine whether apparent insight is supported by the data.
The challenge is therefore about how established model families, constraints and competing scientific views can become trainable components of a PINN without reducing them to fixed preprocessing or allowing the neural component to erase their meaning?
Research architecture
The project treats established empirical and ensemble formulations as structured prior components within a differentiable graph. Their predictions, parameters or intermediate states can inform bounded neural pathways, while learned components represent residual behaviour and interactions that the traditional formulations do not capture. Multiple model views preserve diversity rather than allowing a single convenient representation to dominate.
Evaluation separates predictive performance from scientific behaviour. Ablation, constraint testing, controlled synthetic systems and independent validation are used to ask whether the embedded knowledge contributes meaningfully, whether learned refinements remain credible and when the available observations cannot distinguish alternative explanations. Application-specific graph structure, parameterisation and results remain undisclosed while the first manuscript is in preparation.
Intended contribution
This work contributes a general route to auditable scientific PINNs that can inherit the accumulated value of conventional model ensembles while remaining trainable, testable and extensible. The first paper evaluates the architecture through a focused environmental use case (available soil water); that application is evidence for the wider method rather than the public identity of the programme.