Temperature is one of the most important variables in metal additive manufacturing. In laser powder-bed fusion (LPBF), a highly concentrated laser moves across a powder bed, creating a rapidly evolving thermal field. Local temperature gradients influence melt-pool behaviour, solidification, microstructure and defect formation.

The difficulty is that predicting this thermal behaviour for every new component can be computationally expensive.

Traditional numerical methods, including finite-element approaches, can provide detailed thermal predictions but may require substantial computational resources. A recent 2026 study published in npj Artificial Intelligence investigated whether machine learning could provide a faster alternative.

Demir, Zohdi and Gu developed a convolutional neural-network-based surrogate model for LPBF thermal prediction. Their framework represented geometry and laser-scanning information using signed-distance fields, time fields and time-gradient information. According to the researchers, the resulting model could generalise across previously unseen part geometries and toolpath orientations.

The reported computational advantage is particularly significant: the ML surrogate achieved approximately 1000-fold speed improvement compared with finite-element analysis, while retaining high predictive accuracy for the investigated cases.

Why does this matter?

Consider a complex aerospace component containing hundreds or thousands of scan paths. A manufacturing engineer could potentially use a trained surrogate model to estimate thermal behaviour before the component is printed. Regions likely to experience excessive heating or problematic thermal gradients could then be identified during process planning.

The concept moves additive manufacturing toward predictive process control.

Another important development is the availability of high-speed thermal imaging datasets. A 2026 Scientific Data publication introduced high-speed thermal video data obtained during LPBF of copper components on copper-ceramic substrates. Such datasets can provide spatially resolved information about thermal gradients and hot ejecta associated with defect formation.

This creates a potentially powerful research loop:

thermal imaging → data acquisition → physics-informed ML → thermal prediction → process optimisation.

Instead of relying exclusively on post-production inspection, manufacturers could increasingly predict problematic conditions during the manufacturing process itself.

The broader implication is that artificial intelligence may eventually transform additive manufacturing from a process that is monitored after fabrication into one that is continuously predicted during fabrication.

The major research challenge will be generalisation. A model trained on one geometry, alloy or laser system cannot automatically be assumed to work for another. Reliable AM intelligence will therefore require models capable of understanding manufacturing physics rather than simply memorising datasets.

References:

• Demir, K.G., Zohdi, T. & Gu, G.X., npj Artificial Intelligence (2026).

• Scientific Data (2026), high-speed thermal video dataset for LPBF.