March put AI in chemistry through four increasingly unglamorous tests. Can it generate a polymer that exists? Can public funding make the field share infrastructure? Can it recover critical metals from dirty waste? Can scientific evidence survive an integrity check? The answer depends on which kind of checking is happening.
On March 3, 2026, researchers at the Georgia Institute of Technology published POLYT5 in npj Artificial Intelligence. POLYT5 performs inverse polymer design. Users specify thermal, electronic, and physical properties. The model generates polymer structures that fit them. The team used it to design dielectric polymers for high-temperature energy storage. The team then synthesized a top candidate and measured its properties in the lab.
POLYT5 adapts the T5 encoder-decoder architecture for polymer informatics. It uses Pseudo-SELFIES, a notation based on SELFIES. SELFIES provides mathematically guaranteed structural validity. For polymer chains, the team replaced terminal connection points with Astatine placeholders written as At. The model used span masking during pre-training. It learned from more than 100 million known and hypothetical polymer structures across 20 functional groups and 8 heteroatoms.
Earlier approaches could generate invalid strings, strained rings, or monomers that chemists could not make. Researchers compared POLYT5 with random structure generation, SMILES-based autoregressive models, and classical Density Functional Theory calculations. Synthetic accessibility was checked against known laboratory polymers with scores from 2 to 31. The model achieved 100% syntactic and semantic validity. It was fine-tuned on 12,000 experimentally verified polymers. It generated more than 6 million candidates for glass transition temperature, bandgap, dielectric constant, and melt processability. Property and synthetic accessibility filters reduced that set to 18,000 candidates. A top candidate came from solution polycondensation of glutaryl dichloride and 4,4'-diaminodiphenylmethane. Its predicted glass transition temperature was 483 K. The measured value was 411 K. Its experimental melting temperature was 545 K. Its electronic bandgap agreed with DFT calculations. The syntax passed. The thermometer remained unconvinced.
The 72 K difference shows that sequence-only models still struggle with solid-state chain packing and thermal dynamics. POLYT5 currently handles homopolymers only. It does not yet cover random copolymers, block copolymers, or multi-component blends. Synthetic accessibility filters also reduce suggestions of toxic, explosive, or hazardous intermediates. The system includes an agentic natural-language interface for non-expert users, with chemical formatting and safety checks in place. The next steps are multi-block copolymers, thermosetting networks, and direct links to liquid-handling robots. The work could support materials for electric vehicles, pulsed-power capacitors, aerospace insulation, and flexible electronics.
On March 17, 2026, the U.S. Department of Energy issued Request for Applications DE-FOA-0003612 under the White House-led Genesis Mission. The program offers $293.76 million across 21 challenge areas. The areas include scientific foundation models, advanced materials discovery, nuclear physics, energy storage, and industrial manufacturing. The program funds consortia that join DOE national laboratories, academic institutions, and industrial partners.
The framework focuses on open, self-improving scientific foundation models and shared digital infrastructure. It includes the American Science Cloud and the Transformational AI Models Consortium. Project areas require multi-agent AI frameworks for autonomous research planning, microelectronics design, dynamic fluid simulation, and critical mineral separation chemistry. The plan also links national supercomputers such as Summit and Frontier with industrial operational data and university laboratories. Phase I applications must include national laboratories, higher education, and industry.
The baseline is a collection of piecemeal domain-specific tools. The program aims to replace that model with unified scientific foundation models trained across DOE supercomputing centers. Phase I awards range from $500,000 to $750,000. Phase II research center awards can reach $15 million. The program covers 21 challenge areas selected from 26 national science priorities. This is a funding program, so even the shared AI stack must split the check.
The cost-share rules are a real limit. For-profit partners must provide at least 20% cost share for basic and applied research and 50% for commercial demonstrations. Products, chemical processes, or software inventions from the awards must be manufactured substantially in the United States. For chemical engineering, computational chemistry, and materials science, the program changes access to national laboratories and supercomputers while supporting commercial process AI. It also requires data governance, open science data publishing through the American Science Cloud, and safety protocols for models applied to nuclear assets, chemical plants, and critical power grids. Phase I applications and Phase II letters of intent are due on April 28, 2026. Phase I awardees are expected in late summer 2026.
On March 24, 2026, the National Science Foundation announced eight winning teams in the Tech Metal Transformation Challenge under its STRIDE Ventures program. The challenge operates in parallel with Germany's SPRIND Tech Metal Challenge. Each winning consortium receives up to $2 million for a 10-month Stage 1 prototype phase. The teams combine biological, chemical, and physical-mechanical processes to recover critical metals from domestic industrial waste streams.
The chemistry includes selective ligand adsorption, solvent extraction, and bio-hydrometallurgy. AI models act as predictive orchestrators. They model thermodynamic multi-phase equilibria, optimize microfluidic extraction parameters, and help design chelating ligands for copper, cobalt, and rare-earth elements. The program uses an accelerated, milestone-gated structure. It also creates a transatlantic innovation channel for secondary waste sources such as e-waste and mine tailings.
The baseline is the traditional supply chain. It relies on primary mining and energy-intensive high-temperature smelting. It is also concentrated in foreign jurisdictions. Candidate systems are evaluated against primary mining on purity, energy consumption, waste footprint, and extraction efficiency. This is AI applied to separation thermodynamics rather than simple molecular generation. Industrial waste has not agreed to become a clean benchmark.
The 10-month prototype schedule leaves little time for physical scale-up. The systems must resist chemical fouling, complex matrix interference, and degradation when they process unrefined waste feedstocks. The program matters to process engineers, hydrometallurgists, and environmental scientists working on circular economy systems. It also supports domestic supply security and low-emission chemical waste reduction. Stage 1 evaluations are expected in early 2027. They will determine which teams advance to Stage 2 commercial deployment.
On March 4, 2026, the senior co-author and the Research Integrity Office at the University of Connecticut formally requested the retraction of a paper in Environmental Science & Technology. The paper first appeared online on December 10, 2025. It studied fluoro-contained free radicals and polyfluorinated-like molecules from photoaged fluorinated microplastics. An institutional investigation found extensive plagiarized, fabricated, and falsified data produced with generative AI and image or data manipulation tools.
The falsified material included Electron Paramagnetic Resonance spectra, mass spectrometry outputs, Density Functional Theory calculations, and ReaxFF molecular dynamics simulations. Supporting Information figure S20 matched a figure from Waria et al., published in 2009. The case therefore involved several kinds of evidence at once. It was not limited to generated prose or an altered image.
The case is one of the first high-profile chemical journal retractions in which generative AI was used to fabricate entire physical simulation datasets alongside spectroscopy data. Peer review accepted the paper because the fabricated DFT predictions and EPR spectra appeared internally consistent. The University of Connecticut investigation identified the problems. The first author admitted sole responsibility. The data agreed with each other. They were still not real.
The case exposes a serious weakness in traditional scholarly screening. It matters to publishers, peer reviewers, and computational chemists who rely on multi-modal evidence. On March 26, the Council of Science Editors held a webinar with Science Editor-in-Chief Holden Thorp and Retraction Watch Executive Director Ivan Oransky. The discussion addressed publisher policies for AI detection and peer review watermarking. The next steps include possible cryptographic verification and raw data repository requirements from major chemical societies such as ACS and RSC. A dataset can pass a consistency check without passing a reality check.
Questions about this issue or the underlying research
info@cheairesearch.com