A casting plant for heavy-vehicle powertrains was doing what most factories do: making parts, inspecting them, and catching defects after the fact. Rajasekaran and colleagues redirected the question. Instead of asking which parts are defective, they asked which process parameters predict defects before the pour.
The team collected data from two core-making machines — operational parameters, process conditions, maintenance records — and trained machine learning models to identify the variables most strongly associated with casting defects. The result was not a black-box classifier but a map of causal leverage: which temperatures, pressures, timings, and maintenance states push the process toward failure. Production and quality personnel could then intervene upstream, adjusting parameters before defective cores ever reached the casting line.
This is not a story about ML accuracy, though the models performed well. It is a story about the direction of the analytical gaze. Inspection looks backward — it asks what went wrong. Prediction looks forward — it asks what is about to go wrong. The two require entirely different data architectures. Inspection needs sensors at the end of the line. Prediction needs sensors throughout the process, plus the infrastructure to correlate upstream conditions with downstream outcomes across time.
Detection scales linearly: twice as many parts require twice as much inspection. Prevention scales sublinearly: once you identify that a specific temperature excursion causes 30% of defects, fixing it eliminates those defects across all future production. But prevention requires something detection does not — a causal model linking process states to outcomes. The barrier to predictive quality is almost never algorithmic. It is the absence of the instrumented process history needed to learn the upstream-downstream relationship. The bottleneck is not intelligence. It is memory.