6.2 Robust SCF solver: Augmented Roothaan-Hall Algorithm
TURBOMOLE implements the augmented Roothaan-Hall (ARH) algorithm for the robust optimization of Hartree-Fock and Kohn-Sham density matrices. The ARH method reformulates the self-consistent field (SCF) problem as a nonlinear optimization problem and augments the conventional Roothaan-Hall equations with additional constraints that stabilize convergence. By explicitly targeting the ground-state solution, the ARH approach reduces sensitivity to near-degeneracies and improves convergence behavior compared to standard SCF procedures.
The implementation in TURBOMOLE is based on the work of Ziolkowski et al. and Høst et al. [86, 87, 88]. The method is applicable to both Hartree-Fock and Kohn-Sham DFT calculations for unrestricted and closed-shell calculations.
Input Options
The ARH algorithm is activated using the $arh keyword in the input file:
$arh
Activates the ARH method using the CROP (conjugate residual optimization procedure) algorithm. This is the default option.$arh davidson
Activates the ARH method using a Davidson-type solver for the augmented equations. It needs more storage space than the CROP solver but is numerically more stable. This option is recommended if the CROP algorithm converges to a saddle point.$arh diagfreq=<N>
Does every N SCF iterations a diagonalization. This semi-canonicalizes the molecular orbitals which can lead to significant reduction in SCF iterations needed to reach convergence. But it is less strict in its constraints and might lead to a different local minimum or a saddle point of the electronic state.$arh incore
Enables incore storage of the trial vectors for the Davidson-type solver. This reduces I/O but increases the memory (RAM) demands. By default switched off.
Note that the keywords $scforbitalshift and $scfdamp refer to heuristic convergence accelerations/stabilizations for the DIIS algorithm. They are not used with the ARH algorithm.
Practical Usage Notes
The ARH method is particularly useful in situations where conventional SCF approaches, such as simple diagonalization or DIIS acceleration, exhibit slow convergence, oscillatory behavior, or convergence to incorrect stationary points. Typical examples include systems with small HOMO-LUMO gaps, open shell or radical systems, near-degenerate orbital spaces, transition-metal complexes, stretched bonds, or calculations involving large and flexible basis sets.
The default CROP-based ARH variant ($arh) is recommended for most applications due to its robustness and low memory requirements. It is generally effective for routine calculations and medium-sized systems. The Davidson-based variant ($arh davidson) may be advantageous for larger systems or in cases where a more explicit subspace treatment improves convergence, at the cost of increased memory usage.
When ARH is enabled, it is usually unnecessary to combine it with other SCF acceleration techniques. The method can be used from the beginning of the SCF procedure and is designed to provide a stable and reliable path to the ground-state solution, especially in challenging electronic structure problems.
Comparison with DIIS-Based SCF Acceleration
The ARH method provides an alternative to conventional DIIS-based SCF acceleration schemes. While DIIS improves convergence by extrapolating the Fock or error vectors from previous iterations, it remains fundamentally heuristic and may fail or become unstable in the presence of near-degeneracies, shallow energy surfaces, or multiple competing stationary points. In such cases, DIIS can exhibit oscillatory behavior or converge to excited-state or symmetry-broken solutions.
In contrast, the ARH approach is formulated as a ground-state-directed optimization of the density matrix. By augmenting the Roothaan-Hall equations and solving the resulting nonlinear problem explicitly, ARH enforces stationarity conditions more rigorously and reduces reliance on empirical extrapolation. This leads to improved robustness in challenging SCF regimes, particularly for systems with small HOMO-LUMO gaps, strong static correlation effects, or large and flexible orbital spaces.
From a practical perspective, ARH may converge more slowly than DIIS in well-behaved cases but is significantly more reliable when standard SCF acceleration fails. For this reason, ARH is best viewed not as a replacement for DIIS in routine calculations, but as a robust alternative for difficult electronic structure problems where convergence stability and control over the target solution are critical.