https://www.phase-trans.msm.cam.ac.uk/abstracts/neural.review.html
We examine the utility and methodology of neural networks in materials science, particularly for modelling complex, non-linear relationships. Unlike standard linear regression, these models use mathematical transfer functions and hidden units to capture intricate interactions within large datasets, such as those governing steel properties and weld characteristics. It is emphasised that addressing modeling uncertainty and avoiding overfitting are critical for ensuring that predictions remain reliable when extrapolating into regions with sparse data. A key highlight is the transition from treating networks as "black boxes" to utilising them as transparent tools for discovering new materials and physical patterns. Furthermore, the work advocates for rigorous publication standards, including the public dissemination of datasets and model coefficients to permit independent verification. Ultimately, the sources demonstrate how integrating Bayesian frameworks and physical models can enhance the quantitative understanding of sophisticated engineering problems.
