Most physics.chem-ph traffic is not about process analytics. The slice that is matters, because the methods that decide what models, basis sets and potentials are usable downstream usually appear here first. Four preprints from the first three weeks of May 2026 are worth flagging.
Vibrational eigenstates at scale
Larsson, Le Dé and Gamboni (arXiv 2605.00998, J. Phys. Chem. Lett. 2026) report tree tensor network state methods, coupled to density matrix renormalisation, that compute thousands of vibrational eigenstates in full dimensionality, up to the 33-dimensional Eigen cation. The architecture inherits from multilayer MCTDH. For PAT teams the relevance is indirect but real: any quantitative IR or Raman model that benchmarks against ab-initio reference spectra is bottlenecked by how many anharmonic states the reference can resolve.
Polaritonic spectroscopy gets a reliability protocol
Vergauwe, Toppari and Groenhof (arXiv 2605.06002) take aim at a reproducibility hole in vibrational strong-coupling experiments. They show, using a transfer-matrix model of a pseudo-first-order reaction, that introducing reactants changes cavity thickness enough to distort apparent kinetics when monitored at a single UV/Vis wavelength. Their recommendation is to fit asymptotic extinction rather than fix it. The wider point is that cavity-modified chemistry has been short on protocols for separating optical artefacts from genuine rate effects; this is one of the first that travels well to other groups.
Photochemical decay coordinates by unsupervised ML
Banerjee, Díaz Mirón, Rodriguez and Hassanali (arXiv 2605.08381) use Differentiable Information Imbalance on trajectory surface hopping data to rank the nuclear coordinates that drive conical intersection access. Across five molecules the method recovers known mechanistic coordinates and surfaces competing decay channels. For workflows that touch UV-stability prediction or photodegradation of APIs, the angle of attack - ML as a coordinate-discovery tool rather than a regressor - is the part to watch. Our chemometrics 101 primer covers the supervised counterpart.
ML potentials at DFT accuracy for CSP
Midgley and colleagues (arXiv 2605.28905) present CSP-MACE-A, splitting total energy into intramolecular and intermolecular contributions and pairing MACE-POLAR with an XDM-style dispersion learner. On 19 AstraZeneca test compounds the potential matches PBE DFT; on 28 CSP blind-test compounds it tracks B86bPBE-XDM. Reranking with harmonic free energies improves polymorph stability prediction. For pharmaceutical solid-form work, this is the practical end of the deep-learning chemometrics trajectory: enough accuracy to substitute for DFT, fast enough to use.
None of these are validated tools yet. They are signals about where the underlying methodology will be in twelve to twenty-four months, which is the right horizon for groups planning their next calibration or modelling refresh.