Researchers Debate Whether MLIPs Can Outperform Classical Force Fields

On October 28, MLIP researcher Timothy Duignan and KTH professor Erik Lindorff-Larsen held a multi-round technical debate on X over whether machine-learning interatomic potentials (MLIPs) can replace classical force fields (FFs), covering accuracy, data, validation, and scalability.

Confirmed

• Lindorff-Larsen argued that current MLIPs are not yet as accurate as classical force fields, while acknowledging FFs have their own flaws; he stressed that the "bottom-up approach will most likely work," but existing QM training data types are insufficient, and MLIPs have never been validated on problems where force fields have been benchmarked for over 10 years.
• Duignan countered that classical force fields, without extensive experimental fitting, cannot even compute the activity coefficient of sodium chloride in water accurately — evidence of their reliance on empirical parameterization and poor scalability; he also noted that zinc ion simulation is critical to many biological systems, yet tetrahedral vs octahedral zinc coordination still requires manual fitting and remains unsolved, and many systems/properties like potassium channel conductance are extremely hard to get right with sufficient accuracy.
• Duignan said he remains open-minded: classical force fields could also improve rapidly and ultimately outperform MLIPs in many applications.

Why it matters

The debate reflects a genuine split in the computational chemistry community over whether MLIPs can replace classical force fields: proponents point to the empirical nature and scalability bottleneck of FFs, while skeptics highlight MLIPs' lack of long-term benchmark validation and sufficient training data. The two sides are not simply opposed — both acknowledge that each technical route has clear weaknesses — and future progress may hinge on accumulating QM data and building better benchmark systems.

2026-10-28 ~ 2026-10-28 · 5 related posts

• [source] Researchers debate whether MLIPs beat classical force fields for protein simulation — TimothyDuignan · 2026-10-28
• Force field experts debate: zinc coordination fitting and non-scalable empirical parameterization — TimothyDuignan · 2026-10-28
• [source] Classical force fields could still catch up and beat MLIPs, researcher says — TimothyDuignan · 2026-10-28
• [source] Lindorff-Larsen argues current MLIPs lag classical force fields on biomolecular accuracy — LindorffLarsen · 2026-10-28
• MLIP debate: QM data types insufficient, benchmarks missing vs a decade of FF validation — LindorffLarsen · 2026-10-28