Data Quality in the Fitting of Approximate Models: A Computational Chemistry Perspective

Area of Science:

• Computational chemistry
• Quantum chemistry
• Materials science

Background:

• Empirical parametrization is crucial for quantum-chemistry methods like density functional theory (DFT) and machine-learning (ML) models.
• High-quality data is often scarce, leading to the use of low-cost, low-quality data for fitting parameters.

Purpose of the Study:

• To investigate the impact of low-quality data on the parametrization of DFT methods.
• To determine the acceptable proportion of low-quality data in fitting sets for DFT and ML models.

Main Methods:

• Fitting DFT-type methods using multiple G2/97 datasets of varying quality.
• Analyzing the performance of DFT models with different ratios of high-quality to low-quality data.
• Evaluating the effect of data diversification on model accuracy.

Main Results:

• DFT fitting can tolerate a significant proportion of low-quality data due to the physical basis of DFT and limited parameters.
• Adding large amounts of low-quality data to small high-quality sets may not improve performance.
• Modest amounts of low-quality data can enhance performance when high-quality data is limited.

Conclusions:

• Exercise caution when over 50% of a DFT fitting set comprises questionable data (average error > 20 kJ mol⁻¹).
• Consider data transferability principles to ensure fitting set diversity for robust parametrization.
• Low-quality data can be beneficial for DFT and quantum-chemistry ML models under specific conditions.