Original title: Probabilistic Deep Learning Framework for Phase Transformation Forecasting aided by In Situ High temperature Microscopy
Ioannis Kouroudis, Niki Balestrieri, Stefan Rotzsche, Niccolo Radice +2 · September 2026
Heating and cooling metals changes their microscopic structure, which sets how strong they are. Predicting how this structure shifts during complex heating or cooling paths is very difficult. Traditional cooling charts are static and missing for many new materials. Older machine learning methods also split image analysis, time predictions, and structural forecasts into separate steps. To fix this, researchers created a unified deep-learning system using live microscope images taken at high temperatures. First, a model compresses an initial surface image into a compact digital summary. Next, another network combines this summary, the cooling rate, and the entire temperature history to forecast how a specific iron pattern forms over time. A parallel statistical model predicts the final outcome and refines the forecast to give clear probability ranges. Testing on S235 steel across five cooling speeds showed high accuracy.
Imagine you are baking a cake, but the oven temperature changes depending on which rack you use. Normally you set the temperature once and expect the cake to bake evenly. But here, the heat pattern shifts with every decision you make about where to place the cake, so you cannot rely on a single recipe. In this paper, the steel's internal structure depends on the exact temperature path it experiences, and the model must account for that changing uncertainty at every step.
Materials scientists can now predict how a steel's microstructure will evolve under any heating and cooling path, not just standard ones. This means they can design heat treatments for new alloys without waiting for experimental cooling curves, accelerating the development of stronger, more durable materials. The probabilistic forecasts also tell them how confident they should be in the prediction.
The abstract only demonstrates the method on S235 steel across five cooling regimes for Widmanstatten ferrite evolution. It does not address generalizability to other materials, phases, or thermal histories. The framework also relies on in situ high-temperature confocal laser scanning microscopy data, which may not be available for all labs.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Probabilistic Deep Learning Framework for Phase Transformation Forecasting aided by In Situ High temperature Microscopy — Ioannis Kouroudis, Niki Balestrieri, Stefan Rotzsche et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.