CHIPS-TB: Evaluating Tight-Binding Models For Metals, Semiconductors, and Insulators

Abstract

As semiconductor technologies continue to scale down to the nanoscale, the efficient prediction of material properties becomes increasingly critical. The tight-binding (TB) method is a widely used semi-empirical approach that offers a computationally tractable alternative to Density Functional Theory (DFT) for large-scale electronic structure calculations. However, conventional TB models often suffer from limited transferability and lack standardized benchmarking protocols. In this study, we introduce a computational framework (CHIPS-TB) for evaluating and comparing tight-binding parameterizations across diverse material systems relevant to semiconductor design, focusing on properties such as electronic bandgaps, band structures, and bulk modulus. We assess model parameterizations including Density Functional Tight-Binding (DFTB)-based MatSci, PBC, PTBP, SlaKoNet and TB3PY against OptB88vdW, TBmBJ-DFT and experimental reference data from the JARVIS-DFT database for 50+ materials pertinent to semiconductor applications. The CHIPS-TB code will be made publicly available on GitHub and benchmarks will be available on JARVIS-Leaderboard.

Figures (4)

• Figure 1: Workflow for the evaluation of tight-binding models. Structural information and pre-calculated DFT properties are obtained from the JARVIS-DFT database. The structural data serves as input to various electronic structure methods-including VASP (OptB88vdW (OPT)-DFT), TB3PY (tight-binding), and DFTB (tight-binding) to compute key material properties such as bandstructure, density of states, bandgap (E$_g$), and bulk modulus ($K$). These computed properties are compared against DFT references using the $run\_chipstb.py$ script with a configuration file (e.g., $tb\_input.json$). The resulting performance metrics are reported on the JARVIS Leaderboard to evaluate the accuracy of different tight-binding models.
• Figure 2: Bandgap predictions versus experimental values. Parity plots for (a) OPT-DFT, (b) mBJ-DFT, (c) PTBP, (d) TB3PY, and (e) SlaKoNet methods. The dashed line shows perfect prediction; gray bands indicate ±0.5 eV error. mBJ-DFT achieves the best agreement (MAE: 0.46 eV), followed by SlaKoNet (MAE: 0.81 eV). Among TB methods, TB3PY (MAE: 1.11 eV) outperforms PTBP (MAE: 1.33 eV). Here, MAE represents mean absolute error, RMSE root mean square error, R$^2$ the coefficient of determination, r the Pearson correlation coefficient, and n the number of data points.
• Figure 3: Comprehensive performance metrics for bandgap predictions. Radar chart comparing accuracy, precision (R2), percentages within 0.5/1.0 eV thresholds, and data coverage. mBJ-DFT shows best overall performance; SlaKoNet leads among TB methods for experimental validation.
• Figure 5: Comparison of OPT-DFT and TB bandstructures for Si. The left panels, (a) and (c), display the full band structures calculated using the DFTB and TB3PY models, respectively. The right panels, (b) and (d), show the pointwise differences between the OPT-DFT and TB predictions.