Mechanism discrimination versus property optimisation in sequential experimental design: a simulation study in alloy age-hardening

Introduction. In autonomous laboratories, experimental design is commonly directed by acquisition functions that maximise a target property, as in Bayesian optimisation. Such criteria presuppose a model structure and sample where performance is high, where competing models frequently coincide. The underlying mechanism may then remain unidentified, and predictions unreliable outside the sampled region. An alternative, here termed disagreement-targeted design, selects experiments to discriminate among competing hypotheses rather than to optimise performance.

Methods. A set of competing closed-form mechanistic models was maintained, and a posterior over the set was updated after each measurement using a known-noise Gaussian likelihood. Successive experiments were selected at the point of maximum predictive disagreement among the retained models, following the model-discrimination criterion of Box and Hill. The approach was evaluated on a simulated alloy age-hardening problem, in which hardness rises to a peak and then decreases through over-ageing; property maximisation samples near the peak, whereas discriminating saturation from over-ageing requires the low-hardness regime. Beginning from four observations under which three candidate models were statistically indistinguishable, this approach was compared with greedy property maximisation and with Gaussian-process Bayesian optimisation over 250 trials with a twelve-experiment budget.

Results. The approach identified the generating mechanism (posterior above 0.95) in 100 percent of trials, usually after a single experiment, compared with 2 percent for greedy maximisation (median posterior 0.62) and 70 percent for Bayesian optimisation. The corresponding median error in predicting the over-ageing regime was 2 percent, against 10 and 5 percent. The predominant failure of the optimisation criteria was non-identifiability rather than confident misidentification.

Conclusions. Property optimisation and mechanism discrimination constitute distinct objectives, and acquisition functions designed for the former may provide limited information for the latter. Disagreement-targeted design offers a data-efficient and computationally inexpensive basis for mechanism identification in autonomous experimentation, with potential application across materials and manufacturing technologies.