Prediction of creep failure time using machine learning

Plastic deformation exhibits a complex and stochastic behavior with correlations in space and time, and much information is encoded in such correlations. By applying classical statistical methods, it will be tough to exploit all that information. However, dealing with complex multidimensional and correlated data in an automated way is what machine learning excels at.

In this work, we applied a Random Forest regression algorithm that automatically learns and exploits spatio-temporal correlations in deformation activity. With this model, we estimated as a function of time the remaining time to failure of many samples. The model is trained and evaluated with synthetic datasets produced by the stochastic model introduced in this other work. We first addressed whether predicting a material’s upcoming failure is possible. This question is of great importance, with implications even for earthquake predictability. Moreover, we studied how different material properties and conditions can affect failure predictability. For example, what’s the role of the material’s structural heterogeneity, the temperature, the loading conditions, or the material’s sample size? We found that the machine learning model can predict the approach to failure with higher accuracy than other, more classical methods proposed in the scientific literature. In addition, we found that the remaining time to failure is predictable with a different degree of accuracy that depends on the distance to such time, the material properties, and the loading conditions. Among those, the local strength heterogeneity is a crucial parameter in producing safer materials since it leads to a less abrupt approach to failure. Such behavior means that an upcoming failure can be identified easier and with higher anticipation, which is crucial for safety reasons. Lastly, we found that the bigger a material sample is, the more difficult it is to predict failure. Unfortunately, this might be bad news for earthquake predictability.

In summary, our proof-of-concept study suggests the possibility of training machine learning models on synthetic data to learn the statistical signatures of material failure. Furthermore, using synthetic data for model training can help overcome the scarcity of experimental data since, by definition, reaching failure implies the destruction of the object under study. A trained model could then be deployed in real-world settings to find early-warning signals of failure hidden in the deformation activity of a structure or the earth’s crust.