Alloy design by hand is slow in a way that is hard to appreciate until you have watched a team spend eight months narrowing down a candidate list through trial synthesis. The classic approach: domain knowledge narrows the element selection, published phase diagrams constrain the composition window, and a grid of experiments covers the remaining space. You produce 20 to 40 castings, machine test specimens, and run tensile and fatigue characterization over several weeks per batch.
Prediction-driven composition optimization does not eliminate the experimentation. It changes where in the process you spend the most time. Instead of probing the space evenly, you spend your experimental budget in the regions where the model says performance is highest and uncertainty is lowest. This post walks through a concrete example of how that process looks inside Matforgelab.
The Problem Setup
Consider a medium-complexity structural alloy development problem: a precipitation-hardenable Al-Mg-Si-Cu quaternary system, targeting aerospace fastener applications. The specification is two hard constraints and one optimization objective.
Hard constraints: yield strength at room temperature above 480 MPa, and corrosion resistance better than a grade threshold measured by mass loss per unit area in a salt-fog test. Optimization objective: maximize fatigue life at a specific stress amplitude (350 MPa R=-1 fully-reversed bending), as estimated by high-cycle fatigue testing to 10^7 cycles.
The composition ranges are: Mg from 0.6 to 1.2 wt%, Si from 0.5 to 1.1 wt%, Cu from 0.1 to 0.4 wt%, with Al making up the balance. This is a bounded quaternary space. The underlying Al matrix and standard aging heat treatment are fixed; only composition varies.
Starting Dataset and Descriptor Construction
The team came in with 67 prior experimental data points: yield strength and high-cycle fatigue life measurements on various Al-Mg-Si-Cu compositions made over two years of prior development, all processed under the same heat treatment protocol (T6 temper). This is a good starting point. The data are self-consistent because the processing route did not change, which matters for surrogate quality more than most teams appreciate.
Descriptor construction for this system: composition-weighted elemental features (atomic radius, electronegativity, valence electron count, melting point) plus four thermodynamic features computed from CALPHAD: volume fraction of beta-prime precipitate at aging temperature, estimated solvus temperature, Mg2Si formation energy, and Q-phase stability index. These CALPHAD-derived features cost a few minutes of compute per composition but substantially improve yield strength prediction accuracy because they encode the physics of precipitation hardening that raw composition features cannot capture.
Total feature vector: 31 dimensions per composition. The GP kernel is Matern 5/2 on the standardized descriptor space.
Surrogate Quality Check Before Campaign Start
Before running any active learning, we always validate the surrogate on held-out data. Cluster-based leave-one-cluster-out CV (5 clusters, Ward linkage on descriptor space) gives a cross-validated RMSE of 22 MPa on yield strength and 18 percent log-scale error on fatigue life. These are acceptable for ranking purposes, meaning the model will reliably distinguish good from bad candidates even if it cannot predict absolute values with high precision.
This is an important calibration point. A surrogate with 22 MPa RMSE on yield strength is not useful for certifying a specific alloy composition against a 480 MPa threshold. It is useful for ranking 200 candidates and identifying the top 15 for targeted synthesis. The model is a screening tool, not a replacement for characterization. We state this explicitly in every campaign report we produce.
Candidate Generation and Multi-Objective Ranking
We sample 4,000 candidate compositions from the quaternary composition space using a Sobol sequence (quasi-random, better space-filling than pure random sampling). For each candidate, the GP surrogate predicts yield strength and fatigue life along with posterior standard deviations.
Constraint filtering: candidates with predicted yield strength below the 480 MPa threshold minus one standard deviation are removed. This is a conservative filter: we keep candidates that might plausibly exceed the constraint, not just those predicted to exceed it. About 1,800 of the 4,000 candidates survive this cut in our Al-Mg-Si-Cu example.
Multi-objective ranking uses expected hypervolume improvement (EHVI) over the joint (yield strength, fatigue life) objective space. The Pareto front from the existing 67 data points serves as the reference set. EHVI scores are high for candidates that push the Pareto front outward in both objectives simultaneously.
The top 20 candidates by EHVI are inspected manually by an alloy-aware engineer before going to synthesis. This manual gate catches physically unreasonable compositions, such as candidates where the predicted Mg and Si ratio would lead to excess Si beyond what the Mg2Si phase can accommodate at equilibrium. The model does not know this from the data alone; the engineer does. Two to four candidates are typically removed at this stage.
What the Output Looks Like
The platform outputs a ranked candidate table with: predicted yield strength (mean, 5th percentile, 95th percentile), predicted fatigue life (same format), EHVI score, distance to nearest training point (a proxy for extrapolation risk), and a flag for candidates more than 2 standard deviations from the training centroid in descriptor space.
For the Al-Mg-Si-Cu example, the top three candidates after the manual gate are spread across the composition space: one with high Si/Mg ratio near 1.0/0.7 wt%, one with elevated Cu near 0.38 wt%, and one near the center of the space at 0.85/0.85 Mg/Si with low Cu. All three are low-extrapolation-risk (within 1.5 standard deviations of training data centroid) and have tight posterior intervals.
These three go to synthesis for casting and T6 processing. One additional composition near the boundary of the training data is included as an active learning query: it has high EHVI score but high uncertainty, so its measured value will improve the surrogate regardless of the outcome.
What Happened After Synthesis
Two of the three low-uncertainty candidates hit both constraints. The high-EHVI, high-uncertainty candidate underperformed on fatigue life: the model had predicted median 140,000 cycles to failure at 350 MPa; measured value was 98,000 cycles. The measured value was still informative: it narrowed the posterior variance in that region of the composition space by about 40 percent, improving the model for the next round.
This is the normal pattern in predictive campaigns. The model is not perfect. The value is in the ordering: going into the first synthesis round with three high-probability candidates rather than running 20 random compositions and hoping one works. The team reached a specification-compliant alloy in one synthesis round (4 compositions synthesized and tested) where the prior approach had taken three rounds of 12 to 15 compositions each.
Alloy development by prediction-first does not remove uncertainty. It concentrates your experimental effort where the uncertainty has the most value, and avoids spending budget on compositions the model already knows will not meet spec. That is the core of what prediction-driven composition optimization is.