Abstract:Calculating anharmonic vibrational modes of molecules for interpreting experimental spectra is one of the most interesting challenges of contemporary computational chemistry. However, the traditional QM methods are costly for this application. Machine learning techniques have emerged as a powerful tool for substituting the traditional QM methods. Universal interatomic potentials (UIPs) hold a particular promise to deliver accurate results at a fraction of the cost of the traditional QM methods but the performance of UIPs for calculating anharmonic vibrational frequencies remains hitherto unknown. Here we show that despite a known excellent performance of the representative UIP ANI-1ccx for thermochemical properties, it fails for the anharmonic frequencies due to the original unfortunate choice of the activation function. Hence, we recommend evaluating new UIPs on anharmonic frequencies as an additional important quality test. To remedy the shortcomings of ANI-1ccx, we introduce its reformulation ANI-1ccx-gelu with the GELU activation function which is capable of calculating IR anharmonic frequencies with reasonable accuracy (close to B3LYP/6-31G*). We also show that our new UIP can be fine-tuned to obtain very accurate anharmonic frequencies for some specific molecules but more effort is needed to improve the overall quality of UIP and its capability for fine-tuning. The new UIP will be included as part of our universal and updatable AI-enhanced QM methods (UAIQM) platform and is available together with tutorials in open-source MLatom at https://github.com/dralgroup/mlatom. The calculations can also be performed via web browser at https://XACScloud.com.

What problem does this paper attempt to address?

### What problem does this paper attempt to solve? This paper aims to address the challenges faced in calculating anharmonic vibrational frequencies of molecules. Specifically, traditional quantum mechanical methods (such as the coupled-cluster method CCSD(T)) are accurate but computationally expensive, while density functional theory (DFT) methods are less costly but lack accuracy. Machine learning (ML) techniques, especially universal interatomic potential (UIP) models like ANI-1ccx, hold promise for significantly reducing computational costs while maintaining high accuracy. However, existing UIPs perform poorly in calculating anharmonic vibrational frequencies, primarily due to the inappropriate choice of activation functions. ### Main research content: 1. **Evaluating the performance of existing UIPs**: - The authors first tested the performance of ANI-1ccx in calculating anharmonic vibrational frequencies and found significant errors, including unphysical negative frequencies. 2. **Improving UIP**: - To improve the performance of ANI-1ccx, the authors introduced a new activation function, GELU (Gaussian Error Linear Unit), and retrained the model, resulting in a new UIP model, ANI-1ccx-gelu. 3. **Validating the performance of the new UIP**: - The authors validated the performance of ANI-1ccx-gelu by calculating the anharmonic vibrational frequencies of various molecules. The results showed that ANI-1ccx-gelu achieved accuracy comparable to DFT while having much lower computational costs than traditional quantum mechanical methods. 4. **Fine-tuning UIP to improve accuracy for specific molecules**: - The authors further explored fine-tuning the ANI-1ccx-gelu model to improve the accuracy of anharmonic vibrational frequencies for specific molecules. The results indicated that the fine-tuned model could achieve accuracy close to the CCSD(T) level. ### Conclusion: - **Main contributions**: - Introduced a new UIP model, ANI-1ccx-gelu, addressing the deficiencies of the original ANI-1ccx in calculating anharmonic vibrational frequencies. - Validated the efficiency and accuracy of ANI-1ccx-gelu in calculating anharmonic vibrational frequencies. - Proposed a method to further improve the accuracy of anharmonic vibrational frequency calculations for specific molecules by fine-tuning ANI-1ccx-gelu. - **Future work**: - Further optimize ANI-1ccx-gelu to enhance its robustness across different systems. - Explore more datasets and different fine-tuning strategies to achieve higher precision in anharmonic vibrational frequency calculations. In summary, this paper provides a new solution for efficiently and accurately calculating anharmonic vibrational frequencies of molecules by introducing and improving the UIP model.

Towards Accurate and Efficient Anharmonic Vibrational Frequencies with the Universal Interatomic Potential ANI-1ccx-gelu and Its Fine-Tuning

Accurate and Affordable Simulation of Molecular Infrared Spectra with AIQM Models

Combining quantum mechanics and machine-learning calculations for anharmonic corrections to vibrational frequencies

Tailored anharmonic potential energy surfaces for infrared signatures

Benchmarking phonon anharmonicity in machine learning interatomic potentials

AIQM2: Better Reaction Simulations with the 2nd Generation of General-Purpose AI-enhanced Quantum Mechanical Method

Universal and Updatable Artificial Intelligence-Enhanced Quantum Chemical Foundational Models

Efficient Calculation of the Lattice Thermal Conductivity by Atomistic Simulations with Ab Initio Accuracy

Accurate harmonic vibrational frequencies for diatomic molecules via quantum computing

All-in-one foundational models learning across quantum chemical levels

MP2 Is Not Good Enough: Transfer Learning ML Models for Accurate VPT2 Frequencies

Potential energy surfaces and dynamic properties via ab initio composite and density functional approaches

High-accuracy thermodynamic properties to the melting point from ab initio calculations aided by machine-learning potentials

Evaluating the Anharmonicity Contributions to the Molecular Excited State Internal Conversion Rates with Finite Temperature TD-DMRG

Accelerating the Training and Improving the Reliability of Machine-Learned Interatomic Potentials for Strongly Anharmonic Materials through Active Learning

Performance Assessment of Universal Machine Learning Interatomic Potentials: Challenges and Directions for Materials' Surfaces

Artificial intelligence-enhanced quantum chemical method with broad applicability

Resolution-of-identity approach to Hartree–Fock, hybrid density functionals, RPA, MP2 and GW with numeric atom-centered orbital basis functions

ANI-1: A data set of 20M off-equilibrium DFT calculations for organic molecules

Can G4-like Composite Ab Initio Methods Accurately Predict Vibrational Harmonic Frequencies?