Could AI Have Discovered Fatigue? A Historical Thought Experiment on Prediction, Mechanism, and the Scientist’s Role

1. Why Fatigue Is an Instructive Case

In fatigue research, a model that accurately predicts cycles to failure can be extremely useful. Yet predicting what will happen and explaining why it happens are not equivalent scientific achievements. The distinction becomes consequential when a model must guide the next experiment, extrapolate to a different loading history, or determine whether an apparent material response originates from the specimen rather than from the experimental system. Discussions of AI-assisted science sometimes place these different capabilities under the single term “discovery”. The history of fatigue provides a useful case in which to separate them more carefully.

The question considered here is deliberately counterfactual: what would constitute the discovery of repeated-load failure if the relevant physical concept were not already available? Simply asking a present-day large language model (LLM) to imagine itself in an earlier century cannot answer this question. A modern model may already have encountered fatigue terminology, historical accounts, constitutive models, and modern failure theories during training. Preventing such vocabulary from appearing in a prompt does not remove that information from the model parameters. A historical thought experiment can therefore clarify the epistemic requirements of discovery, but it cannot, on its own, establish either the capability or incapability of contemporary AI.

2. The Historical Boundary Is Less Clean Than It First Appears

Galileo’s Two New Sciences (1638) considered the resistance of solids and the fracture of beams [2]. Hooke’s De potentia restitutiva (1678) developed a language for elastic recovery and springing bodies [3], while Coulomb’s essay, presented in 1773 and published in 1776, addressed statical problems involving structural resistance [4]. Together, these works supplied elements of what might retrospectively be regarded as a pre-fatigue mechanical vocabulary: applied force, deformation, recovery, resistance, and fracture.

They should not, however, be interpreted as evidence that early investigators universally assumed that a structural member was indefinitely safe whenever each individual load remained below its one-time breaking load. The simple rule introduced below is therefore used only as an intentionally incomplete mechanical model, not as a claim about a universal historical doctrine.

A more important qualification arises from Tredgold’s Practical Essay on the Strength of Cast Iron, first published in 1822. Tredgold explicitly described the possibility that a wire subjected to repeated flexure could break under a force smaller than that required to break it in a single application [1]. His explanation invoked permanent set and should not be conflated with a modern microscopic account of fatigue. Nevertheless, the observation itself matters: by 1822, fracture associated with repeated loading had already entered the written engineering record.

This distinction is essential. Recognition of a phenomenon, introduction of a stable technical term, development of a systematic experimental programme, and establishment of a successful physical mechanism need not occur at the same historical moment. The later emergence of fatigue science must therefore be understood against this earlier record [5].

For that reason, the relevant counterfactual cannot be formulated as: give an AI all writings published before 1828 and ask whether it invents fatigue. Tredgold’s repeated-flexure observation would already reveal a crucial part of the answer. A cleaner thought experiment would require a deliberately restricted collection of earlier sources, with every inclusion and exclusion declared in advance. A genuine empirical evaluation would require still stronger controls: either verifiable restrictions on training exposure or an analogous physical problem whose solution was demonstrably unavailable during model training.

3. From an Anomalous Fracture to a Provisional Physical Idea

Consider a simplified thought experiment expressed in arbitrary force units. A nominally identical specimen fails under a single applied force of approximately 100, whereas a force of 50 produces no visible permanent change when applied once. An intentionally limited initial rule might therefore be

$$\mathrm{one-time}\mathrm{nonfailure}\mathrm{at}P=50⟹\mathrm{unlimited}\mathrm{safe}\mathrm{repetition}$$

(1)

Now suppose replicated experiments show that specimens subjected repeatedly to $0\to 50\to 0$ loading eventually fail. Any number of repetitions used in such an example would be illustrative rather than a historical measurement. The observation immediately challenges Eq. (1), but it does not yet establish progressive internal material degradation.

Several competing explanations remain possible. Nominally identical specimens may differ. An initial flaw may have been present. A transient overload may have occurred. Friction, misalignment, or another feature of the apparatus may have evolved during testing. Temperature or environmental conditions may have changed. Cycling may have produced heating, or the specimen may have accumulated a permanent set too small to have been noticed initially.

Before introducing a new material mechanism, a careful investigator would attempt to separate these possibilities. Specimens and fixtures could be compared, applied force and temperature monitored, surfaces inspected, and the experiment repeated independently. Interrupted tests could be introduced to determine whether measurable changes develop in the specimen before final fracture. Only after such alternatives had been challenged would it become reasonable to introduce a provisional history-dependent description. One possible abstraction is

$$R_{n+1}=R_n-\Delta R(P_n,{\mathcal{E}}_n),\mathrm{failure}\mathrm{when}{R}_n\le R_{\mathrm{required}}\left(P_n\right),$$

(2)

where $R_n$ denotes an unobserved reserve of resistance after the nth load application, while ${\mathcal{E}}_n$ represents other relevant conditions. Equation (2) should not be interpreted as a constitutive model. It is a hypothesis-generating representation of the idea that present resistance may depend on loading history. The variable $R_n$ need not correspond to a unique measurable scalar, $\Delta R$ need not always be positive, and several distinct physical processes could lead to similar observed failure times. Neither this notation nor the concept of a continuously changing hidden state is attributed here to any seventeenth- or eighteenth-century author.

Its scientific value lies instead in the experiments that it suggests. If previous loading changes the specimen, then two specimens subjected to different loading histories may respond differently when subsequently exposed to the same load. If the apparent degradation originates in the testing apparatus rather than in the specimen, replacing or exchanging specimens and fixtures should alter the outcome differently. Interrupted experiments might reveal an observable change that develops progressively before fracture. Different candidate mechanisms should therefore imply different experimental consequences.

At this point, an anomalous fracture has been converted into a competition among causal hypotheses. A relation between load level and repetitions to failure would already constitute an important empirical result. Fitting such a relation, however, would not by itself identify the physical process responsible for it. Prediction and mechanism remain logically distinct.

4. What Would Count as an AI Discovery?

A meaningful test of AI discovery should be organised around prospective inference rather than retrospective reproduction of known terminology. Information could be released to the system sequentially: (i) a declared collection of early sources on strength and elasticity; (ii) observations from isolated loading and unloading experiments; (iii) replicated failures under repeated loading; and (iv) new experiments withheld until after the system has recorded its predictions.

The historical material would need to be curated jointly by historians of mechanics and fatigue specialists. Sources revealing later terminology or decisive observations would have to be identified in advance. Tredgold’s discussion of repeated flexure provides an obvious leakage check: a system allowed to read that passage should not subsequently receive credit for independently discovering the underlying phenomenon.

The evaluation should then distinguish several capabilities rather than compressing them into a single judgement of “discovery”. Specifically, the system should be required to:

1.

recognise the inconsistency between its initial mechanical rule and replicated experimental observations without being supplied with terms such as “fatigue” or “cyclic damage”;

2.

formulate at least two competing explanations, including plausible alternatives involving the specimen or experimental apparatus;

3.

propose experiments whose possible outcomes would discriminate between those explanations;

4.

state prospective and falsifiable predictions before observing the corresponding results, and revise or abandon its explanation when those predictions fail;

5.

define the domain of its proposed explanation, identify what remains unresolved, and specify which observable evidence would distinguish its physical picture from competing accounts.

Controls would be essential. Human participants given access to comparable curated material should perform the same task, and the resulting outputs should ideally be anonymised and assessed under preregistered criteria. Evaluation should concern the structure of inference and experimental discrimination rather than rhetorical sophistication. An unrestricted contemporary LLM could still provide a useful assisted-reasoning baseline, but it could not establish historically independent reinvention because its training exposure could not be excluded. Newly constructed physical analogue problems, particularly those whose solutions do not appear in the training record, would offer a stronger route to evaluating genuine prospective discovery.

Results should also be reported separately for anomaly recognition, hypothesis generation, experimental design, prospective prediction, and mechanistic discrimination. Reducing these dimensions to a single statement about whether AI or humans are “more creative” would obscure rather than resolve the scientific question.

Caution is warranted in both directions. In a molecular-genetics discovery task, Ding and Li reported limitations in the hypotheses and interpretations produced by one generative-AI system [6]. That result concerns a particular system and task; it does not establish a universal boundary on AI-assisted scientific reasoning. Conversely, LLMs have outperformed participating experts on a particular neuroscience outcome-prediction benchmark [7]. That finding demonstrates substantial predictive capability, but prediction of outcomes is not equivalent to autonomous construction and experimental validation of a physical mechanism. Neither result resolves the proposed fatigue test. Model architecture, tool use, access to experiments, and the ability to obtain new measurements could all change future performance.

5. The Continuing Role of Scientific Judgement

The distinction developed above has a practical consequence for fatigue research. When a fitted model predicts failure reliably, that predictive capability should be used and its domain of validity established. When a physical mechanism is proposed, a different standard should apply. One should ask which competing explanation the mechanism excludes, which new observation it motivates, and what experimental outcome would require it to be revised or abandoned. A physical explanation earns scientific value not because it is expressed elegantly or resembles familiar theory, but because it discriminates among alternatives and remains accountable to experiment.

This perspective also suggests a constructive role for researchers in an increasingly AI-assisted laboratory. Scientific work includes identifying anomalies that deserve explanation, designing measurements capable of separating rival hypotheses, connecting observations across scales without concealing assumptions, and deciding when the resulting account is sufficiently reliable for engineering use. AI may assist with every one of these activities. Nothing in the present argument requires them to remain exclusively human.

The central point is instead methodological: claims of scientific understanding should be judged by evidence appropriate to understanding. If a system can formulate competing physical hypotheses, design decisive experiments, make successful prospective predictions, and revise its mechanism when confronted with contradictory evidence, that achievement should be evaluated empirically rather than excluded by definition. Whether future AI systems will equal or surpass human researchers in any of these activities is therefore itself an experimental question. It should be investigated rather than settled by intuition, rhetoric, or historical thought experiment alone.