Current Progress and the Transition-Metal Gap Recent machine-learned electronic-structure methods have made significant strides, but a gap remains for transition-metal (TM) and strongly correlated chemistry. Microsoft's deep-learned density functional, Skala (and its update Skala-1.1), has surpassed hybrid functionals on main-group chemistry and reached chemical accuracy on small-molecule atomization energies microsoft/skala-1.1 Broadening access to Skala creates a faster path .... However, its authors note that strongly correlated systems remain an open challenge limited by the cost of reference data [2506.14665] Accurate and scalable exchange-correlation ... microsoft/skala-1.1. Other leading approaches show near misses: the Truhlar group's ML26@MN15 achieved an overall mean unsigned error of 0.89 kcal/mol across 185 databases that include TM systems https://pubs.acs.org/jctcce/article/22/19/9901/5435309/ML26-New-Ingredients-for-an-Integral-Features Reinventing Density Functional Theory: Deep Learning with ..., but this was on training databases rather than held-out sets, and it lacked a TM-specific breakdown. Meanwhile, specialist methods targeting TM systems (like CIDER26SS and LH25nP) still report errors in the 2–5 kcal/mol range on unseen complexes Machine-learned exchange-correlation functionals for ... Machine-learned exchange-correlation functionals for ... System-Specific Reparameterization of Density Functionals with ... Local‐Hybrid Functional With a Composite Local Mixing ....

The Reference Data Bottleneck The primary obstacle to reaching the strict 1 kcal/mol mean absolute error (MAE) threshold is the generation of high-quality training labels. For strongly correlated and TM systems, standard hybrid or double-hybrid functionals err by 1–3 kcal/mol or more. Generating ground-truth training data requires highly expensive wavefunction methods like coupled-cluster (CCSD(T)) or DMRG. While the May 2026 calculation of FeMoco's 76-orbital ground state to near chemical accuracy using DMRG proves such data can be generated Classical computational simulation of the FeMo-cofactor ..., building a training set large enough for a method to generalize to held-out TM systems will require massive, compute-heavy effort. Furthermore, the 1 kcal/mol target is extremely stringent because the reference values for TM benchmarks like MOR41 and ROST61 themselves carry uncertainties of roughly 1–2 kcal/mol, placing the target close to the noise floor of the reference data.

Cost Limits and Lenient Resolution Pathways The requirement that the method cost no more than a standard hybrid density-functional calculation is not the binding constraint. Existing ML functionals like Skala operate at meta-GGA cost, which is well below global hybrids Broadening access to Skala creates a faster path ... Broadening access to Skala creates a faster path .... While highly accurate neural-network wavefunctions (like Orbformer) fail this cost criterion An ab initio foundation model of wavefunctions that ... The Inevitable Transition to Machine Learning in Quantum ..., ML parameterizations of DFT or Δ-ML corrections on a hybrid baseline are computationally viable. Additionally, the parameters of the question are permissive: they allow purpose-built specialist models and require success on transition-metal or strongly correlated chemistry, not necessarily broad universal coverage. A breakthrough on a narrow, accepted benchmark could therefore qualify without solving the entire domain.

Key Uncertainties and Timelines A qualifying result is highly plausible over the next few years as labs scale up compute and AI accelerates theoretical code optimization. An early resolution is possible if a team releases a Δ-ML correction on a hybrid baseline or hits the threshold on a narrow, newly accepted benchmark split. Conversely, a strict interpretation of "a benchmark the field accepts," combined with the inherent noise limits of TM reference datasets and the difficulty of true out-of-distribution generalization, could push the result well into the 2030s. Set against related questions, the timeline for major theoretical breakthroughs favors a median in 2029 rather than an earlier crossing driven by incremental improvements alone. My median expectation is February 15, 2029, reflecting the 1–3 years of intensive data generation and architectural refinement likely required to definitively cross the 1 kcal/mol threshold on held-out strongly correlated systems. The possibility of a narrow qualifying result supports a 10% probability by June 1, 2027, and a 25% probability by February 15, 2028. The reference-data and generalization obstacles leave a substantial later tail: I assign a 75% probability by January 1, 2031, and a 90% probability by January 1, 2034.