Fine-Tuning vs. From-Scratch Training

[matsui-naoki]

Hello everyone,
As someone relatively new to MACE, I've encountered an intriguing situation regarding fine-tuning a foundation model and would greatly appreciate any insights or suggestions from this knowledgeable community.

Background:

I've developed a foundation model trained from scratch using DFT results of 100,000 structures obtained from the trajectory of structural relaxation. The results were encouraging, with MAE scores for energy around 20 meV/atom, for force approximately 40 meV/Å, and for stress about 1 meV/Å^3.
I then attempted to fine-tune this model using a dataset of about 1,000 DFT-MD trajectories of a specific compound at 2000 K (NVT ensemble). While fine-tuning did improve the MAE score from 140 to 80 meV/atom on the test data, training a new model from scratch on this smaller dataset yielded a significantly better MAE score of 40 meV/atom. It's noteworthy that both approaches utilized the same architecture and number of parameters.
In attempts to improve the MAE score during fine-tuning, I've experimented below but they did not improve the score.
・Increasing the learning rate to 0.01
・Adjusting E0s to their average
・Reducing the batch size from 10 to 2

Input parameters for fine-tuning:
2024-03-10 13:32:56.104 INFO: Configuration: {'num_workers': 1, 'batch_size': 10, 'valid_batch_size': 10, 'foundation_model': 'fmace1.1', 'restart_latest': False, 'device': 'cuda', 'name': 'fine_tuning', 'train_file': './data/train.xyz', 'test_file': './data/test.xyz', 'valid_fraction': 0.1, 'max_num_epochs': 100, 'patience': 10, 'default_dtype': 'float32', 'test_dir': None, 'statistics_file': None, 'valid_file': None, 'seed': 123, 'log_dir': 'logs', 'model_dir': '.', 'checkpoints_dir': 'checkpoints', 'results_dir': 'results', 'log_level': 'INFO', 'error_table': 'PerAtomMAE', 'model': 'ScaleShiftMACE', 'r_max': 6.0, 'radial_type': 'bessel', 'num_radial_basis': 10, 'num_cutoff_basis': 5, 'interaction': 'RealAgnosticResidualInteractionBlock', 'interaction_first': 'RealAgnosticResidualInteractionBlock', 'max_ell': 3, 'correlation': 3, 'num_interactions': 2, 'MLP_irreps': '16x0e', 'radial_MLP': [64, 64, 64], 'hidden_irreps': None, 'num_channels': 64, 'max_L': 1, 'gate': 'silu', 'scaling': 'rms_forces_scaling', 'avg_num_neighbors': 1, 'compute_avg_num_neighbors': True, 'compute_stress': True, 'compute_forces': True, 'multi_processed_test': True, 'pin_memory': True, 'atomic_numbers': None, 'mean': None, 'std': None, 'E0s': 'average', 'energy_key': 'energy', 'forces_key': 'forces', 'stress_key': 'stress', 'loss': 'weighted', 'energy_weight': 1.0, 'forces_weight': 1.0, 'stress_weight': 50.0, 'swa_energy_weight': 1.0, 'swa_forces_weight': 10.0, 'swa_stress_weight': 50.0, 'huber_delta': 0.01, 'optimizer': 'adam', 'lr': 0.01, 'swa': False, 'swa_lr': 0.001, 'weight_decay': 1e-08, 'amsgrad': True, 'scheduler': 'ReduceLROnPlateau', 'lr_factor': 0.8, 'scheduler_patience': 20, 'lr_scheduler_gamma': 0.9993, 'start_swa': None, 'ema': True, 'ema_decay': 0.995, 'eval_interval': 1, 'keep_checkpoints': True, 'save_cpu': True, 'clip_grad': 100.0, 'config_type_weights': '{"Default":1.0}', 'wandb': False, 'restart_lr': None, 'foundation_model_readout': True}

Input parameters for From-scratch training:
2024-03-11 14:03:26.265 INFO: Configuration: {'num_workers': 1, 'batch_size': 10, 'valid_batch_size': 10, 'foundation_model': None, 'restart_latest': False, 'device': 'cuda', 'name': 'scratch', 'train_file': './data/train.xyz', 'test_file': './data/test.xyz', 'valid_fraction': 0.1, 'max_num_epochs': 200, 'patience': 10, 'default_dtype': 'float32', 'test_dir': None, 'statistics_file': None, 'valid_file': None, 'seed': 123, 'log_dir': 'logs', 'model_dir': '.', 'checkpoints_dir': 'checkpoints', 'results_dir': 'results', 'log_level': 'INFO', 'error_table': 'PerAtomMAE', 'model': 'ScaleShiftMACE', 'r_max': 6.0, 'radial_type': 'bessel', 'num_radial_basis': 10, 'num_cutoff_basis': 5, 'interaction': 'RealAgnosticResidualInteractionBlock', 'interaction_first': 'RealAgnosticResidualInteractionBlock', 'max_ell': 3, 'correlation': 3, 'num_interactions': 2, 'MLP_irreps': '16x0e', 'radial_MLP': [64, 64, 64], 'hidden_irreps': None, 'num_channels': 64, 'max_L': 1, 'gate': 'silu', 'scaling': 'rms_forces_scaling', 'avg_num_neighbors': 1, 'compute_avg_num_neighbors': True, 'compute_stress': True, 'compute_forces': True, 'multi_processed_test': True, 'pin_memory': True, 'atomic_numbers': None, 'mean': None, 'std': None, 'E0s': 'average', 'energy_key': 'energy', 'forces_key': 'forces', 'stress_key': 'stress', 'loss': 'weighted', 'energy_weight': 1.0, 'forces_weight': 1.0, 'stress_weight': 50.0, 'swa_energy_weight': 1.0, 'swa_forces_weight': 10.0, 'swa_stress_weight': 50.0, 'huber_delta': 0.01, 'optimizer': 'adam', 'lr': 0.01, 'swa': False, 'swa_lr': 0.001, 'weight_decay': 1e-08, 'amsgrad': True, 'scheduler': 'ReduceLROnPlateau', 'lr_factor': 0.8, 'scheduler_patience': 20, 'lr_scheduler_gamma': 0.9993, 'start_swa': None, 'ema': True, 'ema_decay': 0.995, 'eval_interval': 1, 'keep_checkpoints': True, 'save_cpu': True, 'clip_grad': 100.0, 'config_type_weights': '{"Default":1.0}', 'wandb': False, 'restart_lr': None, 'foundation_model_readout': False}

Concern:

It appears that fine-tuning may be restricting the model's ability to adapt, leading to underfitting for out-of-domain data. Conversely, training from scratch on the smaller dataset tends to overfit.

Question:

When loading the foundation model for fine-tuning, are certain parameters inherently fixed to their values in the foundation model, potentially limiting flexibility? How might I better balance the trade-offs between overfitting and underfitting in this context?

[ilyes319]

Hey,

Fine tuning is very much an ongoing area of research for us. To help you, I would need to know more about your training set, test set and the way you do the finetuning.

• What is your training set composed of. Why don't you use the foundation models that we provide?
• How is the test set system different from the training set? Are the two evaluated at the same level of theory?
• What script do you use to do the fine tuning?
• Can you share the log files for the from scratch and fine tuning runs.

[matsui-naoki]

Thank you for your reply!

・The training set for fine-tuning consists of structures containing Pb, K, and F elements. I've tested the model with other systems comprising different elements and observed similar outcomes. The motivations for developing a new foundation model are twofold: (i) My work uses different settings from those in the Materials Project, such as PBEsol and different POTCAR files. (ii) I am interested in more complex structures, including those with defects, and thus aimed to train a new foundation model capable of representing defective structures.
・The foundation model's training data encompass 100,000 structures featuring 55 elements, while the fine-tuning dataset includes 2,000 structures with 3 elements. The training and testing data for fine-tuning are derived from the first 1,800 steps and the subsequent 200 steps of structures from DFT-MD trajectories, respectively. The DFT conditions vary slightly between the foundation model training data and the small dataset used for fine-tuning. Specifically, the former dataset was calculated with in the fine k-mesh density but later utilizes rough k-mesh (only gamma-point).
・I utilized scripts combined from the foundations, universal, and multi-GPU branches.
scripts_and_logs.zip

[matsui-naoki]

I tried training under 2-3 different conditions after that but could not improve the score of the fine-tuned model. My foundation model is apparently stuck in local minima, so it may not be suitable for training out-of-domain data such as high-temperature trajectory. BTW, in GNoME, the fine-tuning foundation model seems to be the checkpoint before the scheduler runs. I will try the same method using MACE.

[gabor1]

one thing is that your number of channels is very small, so there is inherently limited flexibility in your models.

[matsui-naoki]

Thank you for your response! For the foundation model, as well as during the fine-tuning and training from scratch on the smaller dataset, I utilized 64 channels. I experimented with varying the channel size between 16 and 128 for training the smaller dataset (which includes 3 elements) from scratch. Through this experimentation, I observed that changes in channel size do not significantly impact the performance score...