The UZH protocol: Separating errors and constructing improved CP2K basis sets and pseudopotentials

Reliable density-functional simulations require numerical settings whose residual errors are smaller than the chemical and materials trends being interpreted. The UZH protocol presents a closed-loop workflow that calibrates molecularly optimized Gaussian basis sets on small molecules, validates these settings in unary-crystal equation-of-state benchmarks, and identifies whether the limiting approximation is the Gaussian basis or the pseudopotential to produce improved CP2K parameter files.

The core objective of this protocol is to decompose practical CP2K error into a Gaussian-basis component and a pseudopotential component by performing three systematic numerical comparisons.

1. Production CP2K-GTH-UZH calculations are compared against SIRIUS calculations using the same Goedecker–Teter–Hutter pseudopotential in a systematic plane-wave representation. This comparison isolates the Gaussian-basis contribution, denoted as ∆basis(V).

2. SIRIUS-GTH-UZH calculations (using the GTH-UZH pseudopotential in a plane-wave representation) are compared against all-electron full-potential linearized augmented-plane-wave SIRIUS references. This comparison isolates the pseudopotential error, denoted as ∆pp(V).

3. The total deviation of the practical CP2K protocol is then decomposed as: ∆tot(V) = ∆basis(V) + ∆pp(V).

The workflow proceeds through a constructive sequence involving molecular calibration and periodic verification to guide targeted parameter revisions.

(This section summarizes the three-step conceptual process described in Section II.)

1. The first step asks whether the Gaussian basis is chemically balanced on small molecules. This molecular stage establishes a chemically balanced basis set by optimizing MOLOPT basis sets against reference data, such as Gaussian 16/def2-QZVP settings.

2. The second step asks how the settings perform in reproducible unary-crystal equation-of-state tests. This validates the molecularly optimized settings in periodic benchmarks.

3. The third step identifies which component should be modified, using a "three-way comparison between production CP2K-GTH-UZH calculations, SIRIUS calculations using the same Goedecker–Teter–Hutter pseudopotential in a systematic plane-wave representation, and allelectron full-potential linearized augmented-plane-wave SIRIUS references."

The protocol employs specific optimization strategies for basis sets and pseudopotentials based on the error diagnosis.

(This section details the methods used for generating and refining the necessary components.)

1. For basis set optimization, MOLOPT targets a balance between accuracy and numerical stability using a target function that includes the condition-number penalty. This prevents apparently improved molecular fit from becoming numerically unstable in condensed-phase GPW calculations.

2. For pseudopotential optimization, the CP2K ATOM code is used to perform refits against an all-electron atomic electronic state, including third-order Douglas–Kroll–Hess (DKH3) calculations and fitting the GTH pseudopotential.

3. The protocol distinguishes between basis-limited noble-gas and heavy-element cases from pseudopotential-limited transition-metal cases, guiding targeted revisions with the CP2K basis and pseudopotential optimizers.

The final output of the UZH protocol is a validated parameter release that is constructive rather than merely retrospective.

(This section describes the resulting product and its interpretation.)

1. The intended product is a validated parameter release, not only a retrospective assessment of numerical errors. This means the protocol does not merely measure or reduce errors a posteriori, but allows turning verification outliers into validated CP2K parameter files for simulations across molecules and condensed phases.

2. The classification in Table II serves as the decision table for the first optimization pass, converting pairwise error maps into a component that should be changed—either the MOLOPT basis, the GTH pseudopotential, or both.

3. The resulting workflow provides a reproducible path from verification data to systematically improvable CP2K simulations across molecules and condensed phases.

The protocol's implications are that it clarifies what constitutes convergence in density functional theory.

(This section discusses the broader impact on CP2K usage.)

1. The protocol clarifies that Increasing the GPW grid cutoff alone cannot remove a Gaussian-basis error, meaning basis set changes are necessary for basis-limited cases.

2. Conversely, if the SIRIUS-GTH-UZH and SIRIUS-FP-LAPW curves agree, the remaining CP2K discrepancy is a basis issue that can be addressed without changing the pseudopotential.

Based on the provided scientific paper, here are the specific improvements that can be made to Artificial Intelligence (AI) systems, along with what those improved systems would be capable of:

The UZH protocol provides a closed-loop methodology for self-correcting and improving computational models (specifically density functional theory/electronic structure calculations) by systematically decomposing and mitigating numerical errors. This framework can be directly applied to AI research in areas where complex, high-dimensional data modeling or physical simulations are required.

Here are the specific improvements:

Improvement: Implementation of a Diagnostic and Constructive self-correction loop for predictive models based on residual error decomposition (analogous to the UZH protocol).

2. Improvement: Development of an AI agent capable of distinguishing between errors arising from different components (e.g., basis set limitations vs. pseudopotential limitations) within a single prediction, allowing for targeted retraining or refinement rather than brute-force parameter tuning.

3. Improvement: Integration of molecularly optimized, numerically stable basis sets and transferrable core representations directly into the AI's underlying simulation layer (analogous to MOLOPT basis sets and GTH pseudopotentials).

4. Improvement: Creation of a verification workflow that compares predictions across different computational formalisms (e.g., comparing CP2K/Quickstep results against SIRIUS plane-wave references) to quantify the dominant source of numerical approximation error.

The resulting improved AI system can perform the following specific tasks:

Predicting molecular properties (e.g., energy, bond lengths, polarizabilities) with significantly higher reliability and reduced uncertainty across diverse chemical spaces (molecules and condensed phases).

2. Identifying the exact numerical approximation limiting a prediction—whether it is due to insufficient basis set flexibility or inaccurate core/pseudopotential representation—allowing the system to autonomously decide whether to refine its basis parameters or its core model parameters.

3. Generating optimized, transferable parameter sets for subsequent simulations, ensuring that the improved models are not just larger but numerically stable and chemically balanced (i.e., maintaining good condition numbers).

4. Creating a closed-loop research system where verification data automatically feeds back into the model improvement process, leading to a continuous, systematic reduction of numerical errors across all tested chemical systems.