Classical chemistry models keep improving; quantum methods still trail on real systems

Memory is the model

In a preprint from Mehardeep Singh, the MEUWE framework drops the standard Markovian approximation - the assumption that a molecule's environment forgets instantly - in favor of fractional memory kernels that capture power-law memory tails in viscous and structured molecular environments. The named applications are dipole dephasing, conical intersections, spin crossover, and radical-pair magnetoreception: ultrafast condensed-phase processes where environmental fluctuation is the physics, not the noise. m:…5455

The reported numbers: electronic coherence lifetime extended 13.40x relative to Markovian models; a vertical excitation mean absolute error of 37.85 meV from coupled-channel Volterra equations; suppression of coherent recoil back-transfer in rhodopsin photoisomerization by 99.7%; and a fractional Grote-Hynes dynamic barrier transmission of κGH = 0.513620. The pattern is that long-tail memory matters for photochemistry and quantum biology, and Markovian models leave real accuracy on the table. m:…5455

A classical framework reaching gold-standard territory

A second preprint, also from Singh, presents a transferable range-separated quasiparticle framework for electronic structure and dispersion calculations, with subtractive CASCI embedding for multireference problems. On molecules with 10 to 24 electrons, it reports an out-of-sample correlation-energy MAE of 5.809 mHa against CCSD(T) - the classical gold standard - positioning it as an option where Hartree-Fock or MP2 approaches fail. m:…7275

Two concrete demonstrations anchor the claim. Subtractive CASCI embedding resolves the Hartree-Fock permanent-dipole anomaly in biradicaloid ozone, predicting a dipole of 0.5526 D. And time-dependent Hartree-Fock with a fractional memory retardation kernel yields binding-curve MAEs of 0.0155, 0.0421, and 0.1336 kcal/mol for He2, (H2)2, and ethylene dimers respectively. Notably, memory kernels show up here too - the retarded kernel handles non-covalent dispersion, the same long-tail physics as in MEUWE. m:…7275

The quantum baseline still favors classical

The third preprint, from Mohsen Doust Mohammadi, sets a rigorous classical baseline for CO adsorption on small transition-metal clusters of copper, silver, gold, palladium, and platinum - exactly the kind of system quantum computing is supposed to eventually win. Site-resolved vibrational shifts are separated by at least 94 cm-1, well above the 15-20 cm-1 uncertainty, giving clean experimental handles for site assignment. m:…1874

The quantum comparison is sobering. Noiseless UCCSD VQE fails to reach chemical accuracy on these complexes, with errors of 2.0 to 4.3 mEh, while DMRG with bond dimensions of 500 to 1000 provides a converged classical baseline for systems with up to 36 active orbitals and 61 electrons. The author's conclusion is procedural: run entanglement screening before attempting quantum advantage in transition-metal catalysis simulations. That is a polite way of saying most of these systems should not be on a quantum computer yet. m:…1874

The boundary moves from both sides

Read together, the three preprints redraw the quantum-classical boundary from both directions. Classical dynamics is getting sharper through memory kernels; classical wavefunction methods still win on the systems where quantum advantage is most hoped for. For R&D teams allocating compute between quantum pilots and better classical models, this week's evidence favors the latter - at least for these chemistries. All three are preprints, so the numbers are reported results, not settled facts. m:…5455 m:…7275 m:…1874

Why memory kernels keep appearing

It is worth noting that memory kernels appear in two of the three papers independently: MEUWE's fractional kernels for condensed-phase dynamics and the fractional memory retardation kernel in time-dependent Hartree-Fock for non-covalent dispersion. Two different problems, same mathematical fix - the environment's history matters. That convergence is my read, not a claim either paper makes; each paper reports its own numbers on its own systems. m:…5455 m:…7275

For teams running photochemistry or dispersion-heavy simulations, the practical suggestion from these papers is concrete: check whether a memoryless approximation is the error source before reaching for bigger basis sets or quantum hardware. The CO-cluster paper makes the parallel point for quantum computing: run entanglement screening first, because noiseless UCCSD VQE could not reach chemical accuracy on systems where DMRG with bond dimensions of 500 to 1000 already converges. m:…5455 m:…1874