AI for Alloy Design & Composition Optimization

Composition-Property Relationships: What AI Actually Learns

Every alloy is a point in a high-dimensional composition space. A structural steel has 15+ elements that matter — C, Mn, Si, Cr, Ni, Mo, V, Nb, Ti, Al, Cu, B, N, S, P — and each affects multiple properties simultaneously. The traditional approach: decades of empirical formulae (carbon equivalent for weldability, Mn/S ratio for hot shortness, Ni equivalent for austenite stability). These work within narrow composition windows but fail at boundaries and interactions.

ML models trained on alloy databases learn these composition-property mappings without assuming linearity or independence. A gradient boosted model on 50,000 steel compositions can capture the Mn-C interaction on hardenability, the Mo-V synergy in temper resistance, and the deleterious Cu-Sn interaction in surface hot shortness — simultaneously.

Open data/alloy-compositions.csv in the code panel. Each row is an alloy with composition (wt%), processing parameters (rolling reduction, cooling rate, tempering temperature), and measured properties (YS, UTS, elongation, Charpy impact, hardness).

Multi-Objective Optimization

Real alloy design is never single-objective. You need YS > 450 MPa AND elongation > 18% AND Charpy at -40°C > 47 J AND cost below ₹55/kg AND weldability (CE < 0.42). These objectives conflict — higher strength usually means lower ductility, more alloying means higher cost.

Pareto Front Discovery

A multi-objective optimizer explores the composition space to find the Pareto front — the set of compositions where no property can be improved without worsening another. Bayesian optimization with Gaussian process surrogates is particularly effective here because:

Objective space:
  maximize: yield_strength, charpy_impact_minus40
  minimize: alloy_cost_per_kg, carbon_equivalent
Constraints:
  elongation >= 18%
  P <= 0.025%, S <= 0.010%
  C <= 0.20% (weldability)
Search space:
  C: [0.05, 0.20], Mn: [0.80, 1.60], Si: [0.15, 0.45]
  Cr: [0, 0.30], Ni: [0, 0.30], Mo: [0, 0.10]
  V: [0, 0.08], Nb: [0, 0.04], Ti: [0, 0.025]

Indian Application: IS 2062 Grade Optimization

IS 2062 E350/E450 structural steels are the backbone of Indian construction and infrastructure. Mills like Tata Steel, JSW, and SAIL produce millions of tonnes annually. A 0.5% reduction in Mn content while maintaining properties saves ₹200-400/tonne — at 2 MT/year production, that is ₹40-80 crore annually.

AI-driven composition optimization at Tata Steel Jamshedpur identified that micro-alloying with 0.02% Nb + 0.01% Ti could replace 0.3% Mn reduction in E350 plates while maintaining YS > 350 MPa and improving Charpy toughness by 15 J. The net cost per tonne dropped despite the higher per-kg cost of Nb and Ti.

CALPHAD + AI: Thermodynamic Priors as Training Data

Pure data-driven models need thousands of compositions with measured properties. CALPHAD (Calculation of Phase Diagrams) tools like Thermo-Calc and JMatPro can generate equilibrium and Scheil solidification predictions for any composition — but they do not predict mechanical properties directly.

The hybrid approach: use CALPHAD to generate features that bridge composition and properties.

Open data/phase-diagram-data.json — it contains computed phase fractions, solidus/liquidus temperatures, and transformation temperatures (Ae1, Ae3, Ms, Bs) for each composition in the alloy database.

Surrogate Model Architecture

Input layer: [composition (15 elements), processing (4 params)]
       ↓
CALPHAD feature generator: [phase fractions at service temp,
  Ae3, Ms, solidification range, liquid fraction at 1500°C]
       ↓
Combined features: [composition + processing + CALPHAD features]
       ↓
ML model (gradient boosted trees): [YS, UTS, elongation, Charpy, hardness]

The CALPHAD features encode physical knowledge — the model learns that martensite fraction drives hardness, that wide solidification range correlates with segregation (and hence property scatter), that Ae3 depression by alloying indicates hardenability. This typically improves R² by 0.05-0.12 compared to composition-only models.

Phase Stability Prediction

Open data/material-properties-database.json — it includes phase constitution data alongside mechanical properties.

Sigma Phase in Stainless Steels

For MIDHANI's superalloy and specialty steel production (primarily for ISRO, DRDO, and nuclear applications), sigma phase precipitation is a critical concern. Sigma phase — an intermetallic of Fe, Cr, and Mo — forms during service at 600-900°C and causes catastrophic embrittlement.

Traditional prediction uses the Schaeffler-DeLong diagram or empirical PREN (Pitting Resistance Equivalent Number). An ML classifier trained on service exposure data predicts sigma formation probability as a function of composition, service temperature, and exposure time:

Features: Cr, Mo, W, Si, Mn, Ni, N, C, service_temp, exposure_hours
Target: sigma_detected (binary)
Model: Random Forest with SMOTE for class imbalance
Accuracy: 94% on MIDHANI validation set (vs 78% for PREN threshold alone)

Tata Steel High-Strength Grades

Tata Steel's Galvano and Celsius grades for automotive applications require dual-phase (ferrite + martensite) or TRIP (retained austenite) microstructures. Phase stability during forming is critical — retained austenite must transform to martensite during stamping (TRIP effect) but not during storage or paint baking.

AI models predict retained austenite stability (Msσ — the stress-assisted Ms temperature) from composition and intercritical annealing parameters. This replaces XRD measurement during production QC — the model predicts retained austenite fraction and stability from process data alone, flagging coils where the austenite is either too stable (no TRIP effect) or too unstable (transforms during blanking).

Key Takeaways