The Catalyst — Apr 28, 2026

The 2026 MRS Spring Meeting in Honolulu is running this week, and the unofficial theme is watching things move: oxygen atoms migrating through catalyst guts, robots rewriting their own code mid-reaction, and chirality reaching into magnetic sublattices that aren't supposed to feel it. Beneath the headline papers, a quieter pattern is consolidating — provenance systems, validated potential libraries, and LLM-driven legacy data extraction are stitching together the infrastructure that turns "AI predicted a catalyst" into "we can reproduce it next Tuesday." This is a rich week, but the signal isn't any single bombshell — it's that the gap between simulation, synthesis, and scale is closing in several places at once.

• Pinax Provenance System (NIMS): End-to-end tracking of AI-driven materials workflows, including model calls and decision points, published April 22 in Science and Technology of Advanced Materials: Methods.
• Materials Potential Library (MatPL) (ChemRxiv preprint, April 22): Curated, transferable machine-learning interatomic potentials aiming to be the PyPI of computational materials science.
• Evonik–Zeopore Isodewaxing Catalysts (Evonik, April 16): Mesoporized zeolite catalysts targeting fuels, lubricants, renewable diesel, and SAF feedstocks.
• Sublime Systems Commercial Cement Plant (Sublime Systems): Electrochemical cement crosses the 100,000-ton commercial threshold — continuous industrial-scale operation, not a pilot.
• CATL × HyperStrong Sodium-Ion Supply Deal (CATL): A three-year, 60 GWh sodium-ion supply agreement for stationary storage — the largest sodium-ion order on record.
• EPA 2026 Interim PFAS Destruction Guidance (US EPA, April 23): New technology-evaluation framework, updated analytical methods, and a shift to annual (rather than triennial) revision.

For decades, the textbook picture of heterogeneous catalysis treated the surface as the action zone and the bulk as structural scaffolding. A result reported April 21 captures, for the first time, oxygen atoms moving through the interior of a working catalyst — not just along its surface — using operando imaging during an actual catalytic cycle.

If bulk oxygen is a participant rather than a spectator, then the rate-limiting step for many oxidation reactions isn't what kinetic models assumed, and catalyst design has been optimizing the wrong variable. Surface area and surface site density have been the engineering levers; bulk oxygen mobility turns out to be a hidden dial nobody was turning. A complementary Nature paper on interface-controlled bulk oxygen spillover argues the metal–support interface actively gates oxygen entry into the bulk, which could explain why ostensibly identical catalysts behave differently with small changes in support termination.

The success path looks like ¹⁸O isotope tracer studies mapping these pathways quantitatively across oxide systems, followed by a wave of papers on bulk-mobility-engineered catalysts for automotive exhaust, partial oxidation, and fuel cell cathodes. The failure path is a beautiful one-off result that doesn't reproduce in other oxide chemistries — watch whether two or three independent groups confirm bulk transport in different host materials within six months.

Single-atom catalysts have been a darling of academic catalysis for a decade and a manufacturing nightmare for just as long. Ruqian Zou (Peking University), Ju Li (MIT), and Jiong Lu (NUS) just published in Nature Synthesis a "click-locking" strategy — an auxiliary molecule that chemically locks the metal atom in place like a molecular seatbelt before the support forms around it — integrated with a robot-driven high-throughput platform that closes the loop between synthesis and performance testing.

The paper's reproducibility data is what matters. ICP-OES confirmed that recommended precursor ratios translated into measured catalyst compositions with minor deviations, and catalysts prepared robotically and manually under identical protocols showed consistent degradation performance. That's the industrial compatibility test, and it passed. If SACs become a manufacturing technology rather than a publication category, atom-efficient catalysis stops being a slide-deck story and starts eating into the market for traditional supported nanoparticle catalysts in fuel cells, water treatment, and selective oxidation.

Watch for licensing discussions with chemical manufacturers — that's the observable signal. If Chinese catalyst manufacturers or BASF's catalyst division announce platform deals within the next two quarters, the technology has crossed over.

Self-driving labs are excellent at parameter optimization and notoriously brittle in practice. A clogged needle, an evaporated solvent, an unexpected precipitate — the script crashes and a chemist cleans up the mess. Lee Cronin's group posted a ChemRxiv preprint showing an SDL that integrates a multimodal vision-language model into the control loop. When a camera detected an imine condensation that had crashed out as un-stirrable sludge, the system recognized the physical state, consulted solvent solubility priors, and autonomously wrote new Python to inject heated DMF and remix.

The system rescued 14 of 16 intentionally triggered physical failures on the fly, transforming a blind execution script into a physically aware agent. If commercial lab automation vendors fold vision-based recovery into their software stacks — and they will, because the demo is too compelling to ignore — uptime stops being a hand-wave on grant applications and becomes a measured KPI. The failure mode is more interesting: an agent that hallucinates a recovery procedure and ruins a $40,000 catalyst sample. The signal to watch is whether vendors ship recovery agents with auditable decision logs, or just with marketing.

A team from South China University of Technology and Beihang University published in ACS Nano a closed-loop discovery workflow on a fully commercial robotic stack. An adaptive-learning genetic algorithm designed an initial 144-catalyst campaign to train a multilayer perceptron surrogate; the GA-MLP loop then proposed and validated 25 additional candidates. The result: four high-performing multi-elemental catalysts and tetracycline degradation efficiency pushed from about 78% to 93% during the 169-experiment campaign.

What separates this from the usual closed-loop demo is the SHAP interpretability layer revealing nonlinear composition–performance contributions and quantitative guidance on metal fraction and promoter loading windows. The AI didn't just find a better catalyst — it explained why it's better. The platform was commercial, not custom-built, which means the barrier to entry just dropped substantially. Expect Chinese environmental chemistry groups to move on this fast; the failure signal would be replication attempts that don't transfer cleanly between automation stacks.

Daniel Ess's group published in ACS Catalysis (April 21) an evaluation of universal models for atoms (UMA) as a DFT surrogate for homogeneous transition-metal catalytic energy landscapes. The question — can a trained ML model estimate structures and energies for catalytic intermediates and transition states well enough to triage mechanistic hypotheses? — gets a qualified yes across a range of metals, oxidation states, and ligand environments.

If MLIP surrogates become the first stage of a tiered funnel (surrogate → targeted DFT → experiment), mechanism work stops bottlenecking on compute and starts bottlenecking on experimental truth — which is where it should be. The caveat is real: transition-metal chemistry is full of spin-state messiness and weird coordination changes designed to humiliate overconfident models. The success signal is independent benchmarks on hard cases (PCET barriers, multireference systems); the failure signal is groups quietly reverting to full DFT after a few embarrassing predictions.

A decade of perovskite research has treated defects as the enemy and passivation as the cure. New work suggests the picture was wrong in a productive direction: certain defect configurations — particularly at grain boundaries — create local electric fields that assist charge separation, acting as built-in highways that funnel electrons and holes apart before they can recombine.

The implication is that blanket passivation has been optimizing the wrong thing — or at least optimizing it without distinguishing which defects to neutralize and which to leave alone. Selective defect engineering could unlock efficiency gains the field thought were tapped out. A separate Nature paper this fortnight on autonomous closed-loop perovskite fabrication treats reproducibility itself as the optimization target, which is arguably a bigger commercial unlock than another efficiency record. Watch whether independent groups confirm the defect-assist mechanism across compositions; that's the observable signal that this reframes a subfield rather than describing one device.

Sodium-ion has been the chemistry everyone likes in principle. On April 27, CATL announced a three-year, 60 GWh sodium-ion supply agreement with HyperStrong for energy storage — the largest sodium-ion order on record per CnEVPost — following CATL's April 21 tech day and the earlier February passenger-vehicle launch with Changan.

Sodium-ion won't beat top lithium chemistries on every metric, and it doesn't need to. For grid storage, cold-weather performance, safety, and reduced lithium dependence can be enough. CATL has emphasized platform compatibility with lithium-ion pack dimensions, which matters because manufacturing transitions succeed when factories don't have to be rebuilt. Alternative battery chemistry now has purchase orders, which is usually when a field stops being speculative and becomes industrial. The signal to watch is fuller technical disclosure on cycle life and cost — that determines whether sodium-ion is a stationary-storage niche or a real lithium competitor.

This week: a robot rewriting Python to rescue a beaker of sludge, oxygen atoms caught wandering through the bowels of a catalyst they were supposed to be sitting on, and a chiral organic cation whispering its handedness into a copper lattice that has no business listening. The reproducibility crisis didn't get solved — it got version-controlled, which is somehow more honest. Back next week.

Forward this to the colleague who still thinks bulk oxygen is just scaffolding.