Machine Learning | Omid Bidar

Towards data-driven boiling models using a mixture-of-experts approach

Sun, 02 Aug 2026 00:00:00 +0000

Paper ID: IHTC18-1050.

Presented at the 18th International Heat Transfer Conference, 2-7 August 2026, Rio de Janeiro, Brazil.

2026 Annual Technical Meeting of CCP/SIG Nuclear Thermal Hydraulics

Tue, 30 Jun 2026 00:00:00 +0000

Investigating Symbolic Regression for Bubble Departure Diameter Prediction in Subcooled Flow Boiling

Wed, 27 May 2026 00:00:00 +0000

Project details

Student: Samuel Williams
Programme: MEng Mechanical Engineering
Project type: Final-year dissertation
Academic year: 2025-26
Supervisor: Dr Omid Bidar

Overview

Accurate prediction of bubble departure diameter is important in heat flux partitioning models for subcooled flow boiling, where relatively small errors can propagate strongly into predicted evaporation and quenching heat fluxes. This project investigated symbolic regression as a route to developing computationally efficient, interpretable, closed-form expressions that bridge mechanistic boiling models and data-driven optimisation.

Samuel first implemented a deterministic mechanistic baseline in MATLAB, based on the Klausner force-balance model with Colombo and Fairweather improvements, and validated it against a multi-source database containing 378 experimental measurements. Symbolic regression was then applied using PySR in two complementary forms: a hybrid correction to the mechanistic baseline and a standalone correlation for bubble departure diameter. Leave-one-dataset-out cross-validation was used to assess generalisation beyond individual experimental datasets.

Key outcomes

The mechanistic baseline achieved mean absolute percentage errors between 21.0% and 32.4% across the validation data.
Both symbolic-regression approaches identified superheat-driven growth, cap condensation and flow-shear inertia as the dominant physical drivers.
The hybrid model reduced log-RMSE by approximately 15% relative to the mechanistic baseline.
The standalone correlation achieved a mean absolute percentage error of 27.8% on held-out vertical-upflow data, improving on the empirical correlations used for comparison.
The work demonstrated the potential of symbolic regression for producing transparent and physically interpretable closure models for subcooled flow boiling.

DAFoam - Field Inversion and Machine Learning

Mon, 27 Jun 2022 00:00:00 +0000

My contribution

DAFoam is an established open-source discrete-adjoint framework built on OpenFOAM and developed by the MDO Lab community. I did not create the overall DAFoam codebase. My contribution was to first implement its Field Inversion and Machine Learning (FIML) capability, enabling adjoint-based field inversion for data-driven augmentation of Reynolds-averaged Navier-Stokes turbulence models.

This work introduced the capability to infer spatial corrective fields from reference data and supported objectives based on full velocity fields, velocity profiles, surface pressure, skin friction and aerodynamic force coefficients. It established DAFoam as an open-source tool for field-inversion-based turbulence-model development and subsequent machine-learning workflows.

The implementation was presented in the paper An open-source adjoint-based field inversion tool for data-driven RANS modelling and has since been extended by the wider DAFoam development community.

View my commit history in DAFoam