High-entropy alloy design using ML and CALPHAD

Navigating the Vast HEA Compositional Space with Machine Learning

High-entropy alloys present a design space exceeding 10⁶ possible compositions, making exhaustive experimental screening impossible. Machine learning frameworks that combine phenomenological features with physics-based descriptors now allow researchers to focus experimental effort on high-potential candidates, improving phase prediction accuracy from approximately 70% to more than 90%.

Active learning frameworks use iterative feedback loops to efficiently explore this space. At each cycle, the model identifies the compositions most likely to yield target properties, submits them for experimental or simulation validation, and updates its predictions — progressively narrowing thousands of candidates to a tractable shortlist of 10–50 alloys.

Hybrid feature engineering is central to this accuracy gain. Combining primary features such as phase field parameters and phase separation percentages with physics-based features including mixing enthalpy thresholds, atomic size mismatch, and electronegativity ratios gives models the chemical context they need. Where experimental data is scarce — a common problem for refractory HEA systems — data augmentation through CALPHAD simulations and first-principles calculations generates synthetic training data to fill the gap.

Proven empirical composition rules underpin the ML models. FCC stabilisation requires high Ni and Co content above 25 at.% with moderate Al below 10 at.%, delivering elongation above 30%. Introducing Al at 10–20 at.% alongside Ti or Nb additions of 2–5 at.% drives BCC/B2 precipitation, increasing strength by 50–100% via coherent precipitates. For eutectic microstructures — which offer the most compelling strength-ductility synergy — balanced FCC and BCC formers with controlled cooling rates can yield ultimate tensile strengths above 1.5 GPa with elongation exceeding 20%.

Two composition families illustrate what is achievable at room temperature. The (CoCuFeNi)₁₀₀₋ₓMₓ system (M = Al, V, Si, Ti; x ≤ 25 at.%) achieves dual FCC phases with a strength-ductility balance by controlling M element partitioning. Al₀.₃CoCrFeNi forms a single FCC phase with excellent work-hardening capacity — elongation above 50% — and moderate yield strength around 500 MPa. For high-temperature service, aluminium-enriched refractory B2 alloys with BCC/B2 dual-phase structures achieve yield strengths above 1.5 GPa with more than 15% ductility at room temperature, while maintaining strength above 600 MPa at 800°C.

Predicting Phase Stability: CALPHAD, Thermodynamic Parameters, and ML Models

CALPHAD (CALculation of PHAse Diagrams) remains the gold standard for predicting equilibrium phases in HEAs and guiding composition selection before any physical alloy is synthesised. The method scans binary phase diagrams across all sub-systems to construct multi-dimensional phase stability maps, and high-throughput automation can screen thousands of compositions using commercial databases such as TCNI and TCHEA to identify single-phase or dual-phase regions.

A critical practical note: CALPHAD frequently over-predicts σ-phase formation, so cross-verification with experimental XRD and SEM characterisation after arc melting is essential for the most critical candidate compositions. Kinetic considerations — not just equilibrium thermodynamics — determine what phases actually form during processing.

ML models trained on more than 1,000 experimental HEA compositions now achieve 85–95% accuracy in predicting FCC, BCC, or multi-phase structures. Gradient boosting and neural networks lead in performance, while transfer learning extends phase prediction capability to under-explored refractory HEA systems where labelled data is limited. Graph-based models that capture element-element interaction networks further improve prediction of complex multi-phase assemblages. According to Nature materials research, this class of physics-informed ML is increasingly standard in accelerated alloy discovery programmes.

Thermomechanical Processing Routes That Unlock Strength-Ductility Synergy

Even the most carefully designed composition will underperform if processing is not optimised — microstructure control through thermomechanical treatment is the decisive lever for achieving strength-ductility synergy in HEAs. Three processing strategies have demonstrated quantified improvements across multiple alloy systems.

Grain Refinement via Severe Plastic Deformation

High-pressure torsion (HPT) produces nanostructures with grain sizes of 50–200 nm, delivering a 4× hardness increase compared to the as-cast condition in Al₀.₃CoCrFeNi. Cold rolling at 40% reduction followed by recrystallisation annealing at 600–1000°C refines grains to below 10 μm, increasing yield strength by 105% while maintaining 29% elongation. Warm rolling at 450°C balances dislocation strengthening with ductility preservation through dynamic recovery mechanisms.

In CoCrFeNi-based HEAs, the contributions to total strength increase from thermomechanical processing have been quantified: grain boundary strengthening accounts for 60–70% of the gain, dislocation strengthening contributes 20–30%, and solid solution strengthening provides a baseline of 10–15%.

“Cold rolling at 40% reduction followed by recrystallisation annealing increases yield strength by 105% while maintaining 29% elongation — with grain boundary strengthening delivering 60–70% of the total strength gain.”

Precipitation Hardening Strategies

Coherent nanoprecipitate design is one of the most effective routes to high strength without sacrificing ductility. L1₂-Ni₃(Al,Ti) precipitates in an FCC matrix, sized 20–50 nm at 30–40 vol.%, are achieved via aging at 700–900°C for 20–100 hours. B2-(Ni,Co)Al precipitates in a BCC matrix improve room-temperature strength to above 1.5 GPa while retaining 15–20% ductility. Acicular B2 precipitation — preferentially nucleated in high-strain regions during thermomechanical treatment — further enhances work hardening.

Optimisation targets for precipitation hardening are well-defined: a lattice misfit of 0.5–0.8% for optimal coherency strain hardening, a precipitate spacing of 50–100 nm for maximum Orowan strengthening, and a volume fraction of 30–50% for balanced strength-ductility.

Additive Manufacturing Integration

Laser metal deposition (LMD) and selective laser melting (SLM) provide rapid solidification rates of 10³–10⁶ K/s, suppressing brittle intermetallic formation and extending solid solution ranges beyond what is achievable by conventional casting. Post-AM thermomechanical treatment — combining powder plasma arc AM with 60% extrusion and annealing — increases yield strength by 105% and hardness by 94% through grain boundary strengthening, as documented by researchers at ScienceDirect. Functionally graded HEAs enabled by AM allow spatial variation of composition and microstructure within a single component, optimising local properties for stress distribution — an approach tracked extensively in the PatSnap innovation intelligence platform.

Performance Benchmarks Across HEA Systems and Processing Routes

Validated property data from recent literature provides a reliable map of what is achievable — and where trade-offs remain unavoidable. The table below summarises five representative alloy systems across the strength-ductility spectrum.

The eutectic AlCrFe₂Ni₂(MoNb)₀.₃ alloy stands out in this landscape: arc melting produces a yield strength of 878 MPa, an ultimate tensile strength of 2,830 MPa, and 43.7% elongation — a combination that places it in the upper-right quadrant of the strength-ductility map where conventional alloys rarely reach. Hierarchical microstructure design combining nano-scale precipitates, sub-micron grains, and macro-scale eutectic lamellae has been shown to achieve ultimate tensile strengths above 2.5 GPa with elongation exceeding 30%, according to research tracked by Wiley engineering publications.

A Five-Step Integrated Design Workflow for HEA Development

The most effective HEA development programmes follow a structured five-step protocol that integrates computational screening, phase prediction, property modelling, experimental validation, and process optimisation. This workflow reduces the number of experimental iterations by 3–5× compared to traditional trial-and-error approaches, as documented across multiple independent research programmes.

Step 1 — Compositional Screening: Use ML models to generate 10–50 candidate compositions targeting specific phase assemblages (FCC, BCC, FCC+BCC, eutectic). Apply empirical rules: δ below 6%, ΔH_mix between −15 and +5 kJ/mol, VEC adjusted for target crystal structure.

Step 2 — Phase Prediction: Run CALPHAD simulations for the temperature range 500–1400°C to map phase evolution. Validate critical compositions (5–10 alloys) with XRD and SEM after arc melting.

Step 3 — Property Modelling: Estimate yield strength using the Hall-Petch framework: σ_y = σ₀ + k_HP·d⁻⁰·⁵ + σ_SS + σ_ppt. The Hall-Petch coefficient for HEAs is 400–600 MPa·μm⁰·⁵; solid solution contributions range 50–200 MPa; precipitation hardening contributes 200–800 MPa.

Step 4 — Experimental Validation: Synthesise the top 3–5 candidates via vacuum arc melting or powder metallurgy. Measure tensile properties, hardness, and microstructure evolution.

Step 5 — Process Optimisation: Apply thermomechanical treatment — cold or warm rolling at 30–60% reduction plus annealing at 600–1000°C. Fine-tune aging parameters (700–800°C, 20–100 hours) for precipitation-hardened variants. The PatSnap R&D intelligence suite supports literature and patent monitoring at each stage of this workflow.

Looking ahead, generative models including variational autoencoders (VAE) and generative adversarial networks (GAN) now propose novel compositions 10× faster than traditional trial-and-error. Multi-objective Bayesian optimisation simultaneously maximises strength, ductility, and cost-effectiveness, while Frontiers in Materials has documented high-throughput CALPHAD as a key enabler of accelerated metallurgy programmes.