The Catalyst — Mar 24, 2026

A phosphine just did a metal's job under light, an LLM guessed inorganic synthesis conditions right on its first try on a held-out test set more than half the time, and a Nature paper solved perovskite's shading problem by embedding a memristor inside the cell itself. The through-line: chemistry's supporting cast — ligands, language models, circuit elements — keeps getting promoted to lead roles, and the old division between "the thing that does the work" and "the thing that helps" is dissolving faster than anyone's org charts can track.

Phosphines have spent decades as chemistry's most reliable stage manager — coordinating metals, tuning selectivity, never taking a bow. Abigail Doyle's group at UCLA just handed one the lead. Published in Nature (DOI: 10.1038/s41586-026-10263-7), the work shows a phosphine generating radical intermediates under UV irradiation and catalyzing a light-driven reaction between amine-containing rings and alkenes — behavior that looks like oxidative addition and radical rebound, except there's no metal anywhere in the flask. Sami Lakhdar at the University of Toulouse called the result "really spectacular," and the practical kicker is that the overall transformation isn't accessible with transition metals at all.

What changes if this generalizes: chiral phosphine photocatalysis could deliver metal-free asymmetric synthesis of molecules that currently require expensive Pd or Ir catalysts. The team is already pursuing stereoselective variants. The signal to watch: a follow-up in JACS or Nature Chemistry demonstrating enantiocontrol. If that lands within six months, expect pharmaceutical process groups to start screening phosphine photocatalysts seriously. If it doesn't, this stays a beautiful mechanistic curiosity.

The materials discovery pipeline has a bottleneck, and it isn't prediction — it's synthesis. Millions of computationally predicted crystal structures sit in databases with no recipe attached. A new paper in ACS Applied Materials & Interfaces shows that off-the-shelf LLMs (GPT-4.1, Gemini 2.0 Flash, Llama 4 Maverick) achieve 53.8% top-1 and 66.8% top-5 accuracy on predicting precursors for held-out inorganic reactions on a held-out test set, with zero task-specific fine-tuning. That's literature recall at scale — and it's right on the first guess more often than an experienced chemist working from memory alone.

What changes: self-driving labs gain a planning layer that proposes which synthesis to attempt first, not just which steps to execute. Pair this with closed-loop platforms like Polybot — which demonstrated inverse design from target color to novel electrochromic polymer in 72 hours — and the synthesis bottleneck could shrink dramatically. Failure mode: if accuracy drops sharply outside the training distribution (novel oxides, nitrides, high-entropy compositions), this is memorization, not reasoning. Independent benchmarks on out-of-distribution chemistries are the test. Operational caveat: cloud-hosted LLM dependence introduces non-scientific failure modes — API rate limits and billing exhaustion (HTTP 429 errors have already been logged in project workflows) can halt planning mid-run. Multi-provider redundancy and local fallbacks aren't optional.

Perovskite solar cells degrade fast under reverse bias — the unavoidable condition when one cell in a series-connected module gets shaded. The standard fix: make the perovskite tougher. Mohammadi et al. in Nature took a completely different path: they embedded a memristor directly into the cell architecture. When reverse bias kicks in, the memristor switches to a low-resistance state and reroutes damaging current around the perovskite layer — a self-resetting circuit breaker built into the semiconductor stack.

What changes: if fabrication complexity stays manageable, this eliminates the need for external bypass diodes and makes perovskite modules viable in real-world partial-shading conditions. What failure looks like: the memristor integration adds processing steps or materials that erode perovskite's cost advantage over silicon. The signal is whether any module manufacturer announces a licensing or co-development deal within the next two quarters. Meanwhile, a complementary arXiv preprint posted the same day as the Nature paper maps reverse-current degradation dynamics with mechanistic precision — local hot spots, ion migration, filamentary shunting — giving module engineers exactly the failure-mode data they need to pair with this device-level fix.

General-purpose chemistry LLMs hallucinate impossible structures when asked to design metal-organic frameworks — they don't encode the topology and connectivity logic that determines whether a pore is actually accessible and stable. A new arXiv preprint (not yet peer-reviewed) introduces a domain-specific language model trained on MOFid representations — compact strings encoding both chemical composition and topology from the Reticular Chemistry Structure Resource database. The model combines a generative GPT architecture with property predictors and reinforcement learning to optimize candidates for targeted pore geometries and gas-adsorption properties.

What changes: if the generated candidates survive experimental synthesis, this shrinks the MOF design search space from hundreds of thousands of possibilities to a manageable shortlist. The model also offers interpretability — you can query which fragments drive target properties, turning suggestions into actionable design rules rather than black-box scores. Failure signal: if synthesized candidates consistently collapse or show properties far from predictions, the topology encoding isn't capturing enough real-world chemistry. Watch for validation attempts from established MOF groups at Northwestern, MIT, or KAUST.

A reaction that takes 15 hours at 160°C in bulk solvent happens in minutes at room temperature inside a sub-femtoliter aerosol droplet. A new JACS paper (March 20, 2026) used Aerosol Optical Tweezers plus real-time Raman spectroscopy to trap individual droplets, control their size, charge, and water content, and watch an esterification reaction unfold molecule by molecule — fixing the poor parameter control that plagued every previous aerosol chemistry study.

What changes: the air–water interface emerges as a genuinely different catalytic environment from bulk solution, with implications for synthetic chemistry, atmospheric science, and microreactor design. Failure mode: if the rate enhancements don't translate to preparatively useful yields at any scale, this remains a mechanistic insight rather than a synthetic tool. The observable signal: adoption of aerosol optical tweezers setups by synthetic chemistry labs (not just atmospheric groups) within the next year.

A cobalt catalyst enriched with twin boundaries and stacking faults — crystallographic defects that create unusually active coordination environments — converts nitrite to ammonia at ampere-level current densities when paired with microwave plasma air activation. The tandem system splits the hard problem (N₂ activation) into two tractable steps: plasma converts air nitrogen into reactive NOₓ intermediates, then defect-rich cobalt reduces them electrochemically.

What changes: ampere-level current densities bring this closer to industrial operating conditions than most academic electrocatalysis. If the system proves durable over thousands of hours, it's a viable distributed ammonia production route powered by renewables. Separately, iron-substituted MoOₓ OER catalysts from KIMS — published as a ChemSusChem cover — reinforce that earth-abundant defect engineering is where near-term electrolyzer cost reductions will come from. Failure signal: if Faradaic efficiency drops at sustained operation or the plasma step's energy cost negates the electrochemical savings, the system economics don't close.

CSIRO and the University of Melbourne built the first proof-of-concept quantum battery — a device where entangled light-absorbing molecules undergo collective "superabsorption," charging coherently so that charging time decreases as the battery scales up. Published in Nature Light: Science & Applications, the work uses ultrafast laser verification of superextensive charging dynamics, a result that flatly contradicts classical battery scaling rules.

What changes: if energy storage time can be extended beyond the current ultrashort window, this is a paradigm shift in how we think about energy storage physics. What failure looks like: storage times remain too short for any practical application, and the result stays in the quantum-optics curiosity drawer. The signal: a follow-up demonstrating storage on microsecond or longer timescales. Until then, file under "physics that earns its own paragraph but not its own budget line."

• Polybot inverse-design platform (UChicago / Argonne / Purdue): Closed-loop AI-robotic system demonstrated autonomous materials discovery from target property to synthesized polymer in 72 hours — the first reported inverse-design cycle for electrochromic materials without human intervention in the loop.
• NSF Energy Storage Engine Phase 2 (Binghamton University consortium): $45M second-phase funding activated for New York's "Battery Belt" — building lab-to-fab pilot lines and startup formation infrastructure for next-generation battery chemistries in Upstate New York.

A phosphine doing a palladium's job under a UV lamp, a memristor playing circuit breaker inside a solar cell, and a quantum battery that charges faster the more molecules you add — chemistry's supporting cast is staging a coup. Somewhere, a gold nanoparticle is evaporating itself into a shape no one predicted, and honestly, same. Stay curious.

If someone you know would enjoy reading about phosphines with ambition, forward this their way.

The White House issued an AI policy proposal and told states to back off — directly relevant to what self-driving labs and AI synthesis tools will be allowed to do next. Read → The Lyceum on AI Policy