A Further Comparison of MPS and TTNS for Nonadiabatic Dynamics of Exciton Dissociation

Tensor networks, such as matrix product states (MPS) and tree tensor network states (TTNS), are powerful ansätze for simulating quantum dynamics. While both ansätze are theoretically exact in the limit of large bond dimensions, [J. Chem. Theory Comput. 2024, 20, 8767-8781] reported a non-negligible discrepancy in its calculations for exciton dissociation. To resolve this inconsistency, we conduct a systematic comparison using Renormalizer, a unified software framework for MPS and TTNS. By revisiting the benchmark P3HT:PCBM heterojunction model, we show that the observed discrepancies arise primarily from insufficient bond dimensions. By increasing bond dimensions, we reduce the relative difference in occupancy for weakly populated electronic states from up to 60% towards the end of the simulation to less than 10% and the absolute difference from 0.05 to 0.005. We also discuss the impact of tensor network structures on accuracy and efficiency, with the difference further reduced by an optimized TTNS structure. Our results confirm that both methods converge to numerically exact solutions when bond dimensions are adequately scaled. This work not only validates the reliability of both methods but also provides high-accuracy benchmark data for future developments in quantum dynamics simulations.

💡 Research Summary

This paper presents a systematic investigation of the apparent discrepancy between matrix product states (MPS) and tree tensor network states (TTNS) when applied to the non‑adiabatic dynamics of exciton dissociation at a P3HT:PCBM heterojunction. The authors employ the unified Renormalizer software package, which implements both MPS and TTNS within the same time‑dependent variational principle (TD‑VP) framework using a projector‑splitting (PS) integrator. By doing so, they eliminate differences arising from disparate code bases and focus solely on the intrinsic properties of the two tensor‑network ansätze.

The physical model revisited in this work is a reduced representation of the P3HT:PCBM interface: a linear chain of 13 oligothiophene (OT) molecules, each bearing a local excitation (LE) and a charge‑separated (CS) electronic state, coupled to a fullerene cluster. The system includes 26 electronic states and 113 vibrational modes (8 local OT modes per molecule, 8 fullerene modes, and a single intermolecular mode R that mediates the LE₁↔CS₁ transition). The Hamiltonian is partitioned into electronic, vibrational, and electron‑phonon coupling parts, with parameters taken from earlier benchmark studies.

Key methodological points include:

1. Unified TD‑VP‑PS implementation – Both MPS (equivalent to TD‑DMRG) and TTNS (a variant of ML‑MCTDH) are propagated with identical time‑step control and error handling.
2. Fixed bond dimension (M) across all bonds – To isolate the effect of bond dimension, the authors use a uniform M for every edge in the network, expanding the initial product state by adding a small set of Krylov vectors with weight 10⁻¹⁰.
3. Entanglement‑entropy guided analysis – The von Neumann entropy S for each cut is computed from the singular values of the canonical tensors, providing a quantitative estimate of the minimal M required (M ≥ e^S).

The convergence study varies M from 64 up to 512. Results show that the previously reported up‑to‑60 % relative error in the population of weakly occupied electronic states (e.g., higher‑order CS states) is dramatically reduced when M is increased. At M = 256 the relative error falls below 10 % and the absolute population difference shrinks from 0.05 to ≈0.005. Energy conservation and total norm remain within 10⁻⁸, confirming numerical stability.

A central contribution is the design of an optimized TTNS topology. Guided by the entanglement profile, the authors allocate larger bond dimensions to edges surrounding the strongly coupled LE₁↔CS₁ transition and the R mode, while using smaller dimensions for peripheral OT vibrations. This non‑uniform tree reduces computational cost by roughly 30 % compared with a naïve balanced 3‑leg TTNS, yet yields even lower errors than the standard MPS at the same nominal M.

Performance metrics reveal that, while MPS enjoys O(N M²) memory scaling due to its linear chain, the optimized TTNS achieves comparable wall‑clock times because the reduced bond dimensions on many edges offset the nominal O(N M³) scaling of a generic tree. Consequently, both methods converge to the same numerically exact solution when the bond dimension is sufficiently large, and the choice between them can be guided by the specific entanglement structure of the problem rather than by concerns over intrinsic accuracy.

In conclusion, the paper resolves the earlier inconsistency by demonstrating that insufficient bond dimensions, not a fundamental limitation of either ansatz, caused the observed discrepancy. It validates the reliability of both MPS and TTNS for high‑precision quantum dynamics, provides a practical protocol for assessing required bond dimensions via entanglement entropy, and introduces an efficient TTNS architecture tailored to exciton‑dissociation dynamics. These insights furnish the community with robust benchmark data and methodological guidance for future simulations of complex organic photovoltaic materials, charge‑transfer processes, and other non‑adiabatic phenomena.