#32 · ML / AI methods

Port-Hamiltonian and scale-free models test whether EEG and circadian dynamics are…

Across four preprints, the dominant thread is evaluating structured surrogate models of biological dynamics by whether they recover interpretable physical or statistical properties, not only predictive accuracy. Two port-Hamiltonian GNN digital twins, one for motor-imagery EEG and one for circadian multi-omics in mouse liver, are reported as stable and forecasting-capable. The EEG model reproduces avalanche criticality, whereas the liver model's network topology recovery is at chance. An RG-Flow transformer matches a vanilla transformer on sleep staging (77.3% vs 77.0%) yet recovers the spectral exponent β out-of-sample. A separate information-theoretic bound states that prediction r² is capped by the variance captured by the projection basis, with graph wavelets outperforming PCA. What is new is the explicit diagnostic framing: accuracy alone does not establish mechanistic fidelity.