It’s been a couple of weeks since I shipped Ferrumizer. Building a differentiable physics engine that simulates industrial steel hardening forced me to bridge my degree in Metallurgical Engineering with modern software infrastructure.
I didn’t want to just build another dashboard; I wanted a tool that respects the gritty reality of manufacturing and doesn’t take mathematical shortcuts.
Swipe through the carousel to see how the different engines actually work under the hood:
1. The Virtual Furnace (The Forward Model)
Instead of assuming an impossible “instantaneous quench,” the solver models finite-rate spatial quenching. By integrating Scheil-JMAK C-curves, it calculates exactly where austenite converts to soft pearlite or bainite instead of hard martensite.
2. The CCT Diagram (The Metallurgist’s View)
The engine generates Continuous Cooling Transformation diagrams, plotting the actual cooling curves at the surface, mid-radius, and core of the part against the transformation noses.
3. The Bayesian Calibrator (The Inverse Solver)
The engine can run backwards. It runs NUTS MCMC against measured shop-floor hardness data to reverse-engineer the furnace’s actual performance. It is guarded by a 200-simulation SBC gate, so if numerical limits are violated, it fails loudly instead of silently spitting out an overconfident, wrong answer.
4. Defensive Log Ingestion (The Blue Collar Reality)
Real factory PLCs output brutal, noisy logs. The ingestion layer runs Ramer-Douglas-Peucker line simplification to compress wandering sensor data into clean schedule knots, preserving the actual heating ramps instead of mutilating them into flat staircases.
The entire stack is open source for anyone who wants to dig into the PDEs or the JAX autodiff wiring: GitHub - Idle0x/ferrumizer · GitHub
Watch: Physics Engine for Industrial Heat Treatment Honesty and Calibration | Fawaz Azeez posted on the topic | LinkedIn
#ScientificML #Metallurgy #PhysicsInformedAI #BayesianInference #SoftwareEngineering