# Graph Contrastive Learning for Selective Critical Metal Recovery from E-Waste

Electronic and battery waste is one of the richest secondary sources of critical metals we have — copper, nickel, cobalt, zinc, gold, and lithium all end up in the same scrap stream. The difficulty is that they also end up in the same leachate. Once that scrap is dissolved, recovering any single metal means pulling it out selectively while leaving four or five competing ions behind in solution.

That selectivity is controlled almost entirely by one decision: which extractant molecule — the organic compound used in solvent extraction — you reach for. Today that decision is still made by empirical trial-and-error and chemist intuition.

Where the Existing Work Stops Short

Machine learning is already applied heavily in this space, but almost entirely to the wrong half of the problem. The bulk of it optimizes process conditions — pH, temperature, residence time, reagent concentration — for an extraction chemistry that has already been chosen. The chemistry itself stays a fixed input.

Separately, graph-based models have shown they can predict how strongly a molecule binds a given metal ion. Two limitations keep that work from transferring to recycling:

• Metal ions are treated as isolated nodes. Existing metal-ligand GNNs (the D-MPNN formulation, for instance) place the metal in the graph with no explicit coordination bonds to the ligand's donor atoms. The coordination environment — the chemistry that actually drives selectivity — is left unmodeled.
• Evaluation happens on clean, single-metal-family systems. Rare-earth separation on well-defined synthetic solutions is the common benchmark. Real e-waste leachate is a substantially harder, messier setting.

The gap is a question of framing. Prior models answer "does this bind?" In a competing-ion leachate, the question a recycler actually needs answered is "does this bind this metal instead of the other four in the same solution?"

Proposed Architecture: A Two-Tower Contrastive GNN

The core idea is to represent a molecule and a metal ion as two vectors — embeddings — living in the same numerical space, then train the model to place compatible pairs close together and competing pairs far apart.

The architecture has two towers:

• Extractant tower. Extractant molecules are encoded as molecular graphs, atoms as nodes and bonds as edges, with the metal ion's coordination environment explicitly represented rather than left as a bare isolated node.
• Metal tower. Target and competing metal ions are encoded as feature vectors through a separate encoder.

A contrastive objective ties them together. For a given metal ion, the embedding of its correct extractant is pulled close; the embeddings of extractants suited to competing metals are pushed away. The similarity between the two towers becomes a direct selectivity score rather than a raw binding affinity.

This is the substantive shift: the model is optimized end-to-end for discrimination between a target metal and its real-world competitors, which is the property that matters industrially — not binding strength measured one pair at a time.

Dataset Strategy

Pretraining starts on NIST SRD 46 (Critically Selected Stability Constants of Metal Complexes), a public curated dataset of roughly 49,000 stability constants spanning 6,173 ligands and 216 metal ions. It is enough to pretrain and sanity-check the base GNN before the selectivity objective and experimental data are introduced.

It is also not sufficient on its own, and it's worth being explicit about why. SRD 46's last major update was around 2004, and it skews toward the simple mono- and bidentate ligands of classical coordination chemistry. It likely under-represents the extractants that hydrometallurgy actually runs on today — D2EHPA, the Cyanex series, the LIX series, glycine-based systems.

So the dataset is supplemented from two directions: compiled distribution-coefficient data from the solvent-extraction and hydrometallurgy literature, and lab-measured data generated in-house. Treating that supplementation as an early, explicit milestone — rather than an afterthought once training stalls — is part of the plan.

A Closed Predict-Test-Refine Loop

The work is designed as an integrated computational and experimental pipeline, not a purely computational study:

1. Curate the candidate extractant pool and the ligand-metal training data.
2. Build molecular graph and metal-ion representations.
3. Pretrain and train the two-tower contrastive GNN.
4. Screen and rank synthesizable extractants by predicted selectivity for a difficult metal pair — Cu vs. Ni, or Co vs. Ni.
5. Validate the top- and bottom-ranked predictions with real solvent-extraction experiments on leachate generated from actual waste.
6. Fine-tune on those experimental results, closing the loop.

Validating on real secondary-waste leachate rather than synthetic solution is deliberate, and it's the part most likely to be humbling. It is also the only version of the experiment whose result means anything to a recycler.

The payoff, if the loop holds, is narrowing thousands of untested candidate molecules down to the handful worth putting on a bench — and having each round of bench results make the next round of predictions sharper.

Open Questions

Three questions remain genuinely open at this stage: which metal pair to commit to as the primary target, whether PCB scrap or spent battery material is the more practical feedstock for repeated leachate generation, and how many extractant candidates can realistically be tested given lab time and reagent budget. The answers shape the scope considerably, which is why they're being settled before training begins rather than after.

---

This post summarizes a research proposal co-authored with Harshavardhan Gowda, Department of Metallurgical and Materials Engineering, National Institute of Technology Karnataka, Surathkal.