Machine-Learning-Guided Calibration of Lobe-Specific Breathing Profiles in the TIDAL Model

• Divyansh Jain, Chemical Engineering, University of Delaware

• Catherine Fromen, Chemical and Biomolecular Engineering, University of Delaware

Machine-Learning-Guided Calibration of Lobe-Specific Breathing Profiles in the TIDAL Model

Machine-Learning-Guided Calibration of Lobe-Specific Breathing Profiles in the TIDAL Model

Divyansh Jain¹, Dominic Hoffman¹, Logan Whitesel², John Thomas¹, Catherine Fromen^1,2

1 – Chemical and Biomolecular Engineering, University of Delaware, 2 – Biomedical Engineering, University of Delaware

Abstract

Inhaled aerosols can deliver medication directly to the lungs, potentially concentrating treatment at the disease site while reducing exposure elsewhere in the body. Their effectiveness depends on where particles deposit, which changes with breathing mechanics and asymmetric filling of the five lobes. Reliable testing therefore requires an in vitro lung model that reproduces clinically measured lobe-specific breathing patterns.

The Total Inhalable Deposition in an Actuated Lung (TIDAL) model uses five independently actuated lobes to simulate breathing. Its original fixed motor settings approximated healthy adult male breathing but did not precisely match the lobe-specific reference data; refinement required time-intensive manual tuning. The goal of this project was to automate this process so that clinically accurate breathing profiles could be generated rapidly from four-dimensional magnetic resonance imaging (4D MRI) data.

A closed-loop machine-learning pipeline used Bayesian optimization to propose motor and waveform settings, test each candidate in a single lobe study, and compare measured airflow with the clinical reference profile. Results from completed trials guided progressively more accurate configurations. The optimized lobe profiles were then run simultaneously to determine whether independently calibrated motions combined into a coherent whole-lung breath.

Preliminary testing showed closer agreement with the clinical reference than the original settings, and all five optimized profiles operated together while preserving the intended lobe-specific asymmetry. By replacing manual calibration, the workflow provides a faster, reproducible method for generating clinically informed TIDAL profiles and studying where inhaled medicines deposit throughout the lungs.