Irreversible processes are usually modelled by stepping forward in time, but inferring past states from partial observations is a separate challenge. This paper proposes a probabilistic framework that treats the entire history of a system as the unknown, ranking all admissible histories with a Gibbs-type measure exp(-J/Theta). Here J encodes both constitutive physics and observational constraints, while Theta controls how much epistemic uncertainty remains. Maximising this posterior is equivalent to minimising J over the whole history space, which unifies forward-in-time evolution and inverse inference in a single formulation.
The framework is specialised to generalized standard materials (GSM), where J is built from incremental energy–dissipation functionals. The analysis reveals an important distinction: the history obtained by simultaneous minimisation over all admissible histories generally differs from the one produced by sequential GSM updates. The causal, sequentially generated history incurs a higher history-level cost, indicating that enforcing causality carries an additional penalty at the history level. This result is illustrated across seven representative examples, including viscoelasticity, viscoplasticity, rate-independent plasticity, gradient damage, phase transformations, and coupled thermoelectric evolution.
The paper also demonstrates the finite-Theta formulation on an inverse problem with a non-convex free energy. Using incomplete initial-state information, the method infers unobserved intermediate states and quantifies uncertainty over admissible histories. As Theta decreases, the posterior concentrates towards minimum-J histories, showing how the framework can interpolate between broad uncertainty and sharp, deterministic predictions.