Challenges for Foundation MLIPs
==Foundation MLIPs still face several unresolved challenges== spanning what they can accurately describe , how they scale computationally , how they are trained/evaluated , and how reliably they can be deployed . [:c...
==Foundation MLIPs still face several unresolved challenges== spanning what they can accurately describe , how they scale computationally , how they are trained/evaluated , and how reliably they can be deployed . [:cite[1]{ln=1}], [:cite[1]{ln=2}], [:cite[1]{ln=3}] Challenges that remain for foundation MLIPs 1. Physical fidelity: long range interactions Many MLIPs rely on locality assumptions (atomic environments within a cutoff), which implies long range interactions are neglected , including long range electrostatics (and analogously, missing long range correlations like dispersion). [:cite[2]{ln=3}], [:cite[2]{ln=4}] More specifically, foundation MLIPs’ fixed cutoff radius and limited information diffusion through message passing restrict inclusion of long range electrostatics and dispersion, which are crucial for large assemblies (e.g., MOFs, biomacromolecules). [:cite[3]{ln=1}], [:cite[3]{ln=2}], [:cite[3]{ln=3}] The field still lacks consensus on the best way to incorporate long range effects (explicit physics vs. learning from data), and generating accurate reference data for the fully data driven option is expected to be difficult. [:cite[4]{ln=2}], [:cite[4]{ln=3}], [:cite[4]{ln=4}] Incorporating beyond local information can also break strictly linear scaling (i.e., computational complexity may no longer scale strictly linearly). [:cite[5]{ln=8}] 2. System description limits: charge, spin, and magnetism Many MLIPs neglect total charge and spin multiplicity , and even when spin multiplicity is included, it can be insufficient for magnetic phenomena that require local magnetic moments (e.g., antiferromagnetism). [:cite[6]{ln=1}], [:cite[6]{ln=2}], [:cite[6]{ln=3}] Even with recent progress using total charge and spin as inputs, including local magnetic moments and extrapolating to unseen spin configurations and charge states remains a challenge. [:cite[7]{ln=4}] 3. Scalability & efficiency for higher order information (Hessians) Many applications (e.g., transition state searches) require Hessians , and incorporating Hessian information can be important for successful TS searches and can improve vibrational frequency predictions. [:cite[8]{ln=2}], [:cite[8]{ln=3}], [:cite[8]{ln=4}], [:cite[8]{ln=5}] However, computing Hessians for large MLIPs scales quadratically with atom count and requires significant memory , so training on Hessians and obtaining Hessians efficiently at inference remain open challenges. [:cite[8]{ln=6}], [:cite[8]{ln=7}] 4. Training data constraints, dataset scaling cost, and data quality control Foundation MLIPs’ applicability is ultimately constrained by their training data (e.g., models trained largely on bulk material DFT data can have difficulty predicting surface energies; adsorption energies may require additional fine tuning). [:cite[9]{ln=1}], [:cite[9]{ln=2}], [:cite[9]{ln=3}] Scaling training to ever larger, diverse datasets is computationally expensive , and the paper notes the need for a clearer understanding of which data is important for training. [:cite[10]{ln=2}], [:cite[10]{ln=3}], [:cite[10]{ln=4}], [:cite[10]{ln=5}] As datasets grow, erroneous/noisy data becomes more likely and can limit learning; the paper cites evidence of numerical noise in DFT forces and calls for minimal datasets plus fully automated mechanisms to ensure sufficient data quality. [:cite[11]{ln=2}], [:cite[11]{ln=3}], [:cite[11]{ln=4}], [:cite[11]{ln=5}] 5. Evaluation: benchmarks aligned with real applications Evaluations often focus on energy/force errors, but small force errors can still correspond to failures in molecular dynamics , motivating the need for benchmarks closer to actual applications. [:cite[12]{ln=1}], [:cite[12]{ln=2}] The paper also highlights the need for training paradigms that adapt models directly to tasks like MD, rather than using only energies and forces. [:cite[12]{ln=4}] 6. Uncertainty quantification (UQ) and calibration Current foundation models largely lack integrated mechanisms for uncertainty quantification , so users often must use post hoc approaches. [:cite[13]{ln=1}], [:cite[13]{ln=2}], [:cite[13]{ln=3}] For practical use, models must have well calibrated uncertainty estimates , and designing UQ that is both mathematically rigorous and computationally efficient remains an open challenge. [:cite[13]{ln=5}], [:cite[13]{ln=6}] The paper also warns that uncertainty quality must be evaluated and notes reports of significantly underestimated uncertainties . [:cite[14]{ln=4}], [:cite[14]{ln=5}] 7. Model maintenance: fine tuning risks catastrophic forgetting Fine tuning a generalist model to add new knowledge increases the risk of catastrophic forgetting (performance deterioration in previously mastered areas). [:cite[15]{ln=1}], [:cite[15]{ln=2}] Preventing forgetting while keeping enough flexibility to incorporate new information is framed as the stability–plasticity ...