Image-Based Scientific Machine Learning for Theories of Biological Dynamics Across Scales – Mechanical Inference from Biological Images

Mechanical inference seeks to extract quantitative information about forces, stresses, mechanical parameters and physical models directly from biological images. In epithelial and multicellular systems, this inverse problem can be formulated at several levels: quasistatic tension/pressure inference from cell geometry; Bayesian estimation under incomplete determinacy; inference from dynamic image sequences; 3D force inference from reconstructed cellular interfaces; tissue-level stress inference; and, increasingly, direct inference of model parameters or governing laws using differentiable simulation, stochastic inference, continuum mechanics and machine learning.

This meeting is conceived as a technical workshop and working meeting, not as a school or a sequence of conventional conference talks. The aim is to put closely related methods side by side, make their assumptions explicit, work through selected derivations or examples at the blackboard, and confront them with common image datasets. Formats would mix concise scientific presentations, presenter-led or collective blackboard sessions, a small number of live worked examples, and one prepared practical comparison exercise.

A recurring theme will be determinacy and falsifiability: which observables constrain which unknowns; how priors and constitutive assumptions close underdetermined problems; how 2D and 3D formulations differ; and which independent experimental measurements can invalidate an inferred force field. A second goal is to connect classical force inference to newer inverse-problem formulations in optimization, stochastic dynamics, continuum/FEM models and differentiable image-based modeling.

The workshop would culminate in a proposal for a community benchmark for mechanical inference from images. Rather than organizing an open-ended hackathon, a small set of datasets would be prepared in advance, including synthetic 3D microscopy generated from several forward mechanical models with known ground truth, together with controlled segmentation/meshing variants and selected experimental datasets. Participants could then compare inference methods and identify where discrepancies arise from the images, geometric reconstruction, model mismatch, or the inference procedure itself.