MatExpert: Decomposing Materials Discovery By Mimicking Human Experts

Authors: Qianggang Ding, Santiago Miret, Bang Liu

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results demonstrate that Mat Expert outperforms state-of-the-art methods in material generation tasks, achieving superior performance across various metrics including validity, distribution, and stability.
Researcher Affiliation Collaboration Qianggang Ding1,2, Santiago Miret3, Bang Liu1,2 1 DIRO & Institut Courtois, Universit e de Montr eal 2 Mila Quebec AI Institute 3 Intel Labs EMAIL EMAIL
Pseudocode No The paper describes a three-stage methodology (retrieval, transition, generation) with detailed textual explanations of each step and the models used. However, it does not present any structured pseudocode or algorithm blocks with numbered or bulleted steps formatted like an algorithm.
Open Source Code Yes Codes are available at: https://github.com/Bang Lab-Ude M-Mila/Mat Expert
Open Datasets Yes 2. We curated a large-scale dataset from the NOMAD Scheidgen et al. (2023) database, which serves as a comprehensive testbed for evaluating Mat Expert across diverse material compositions. We will publicly release the datasets and source codes upon publication.
Dataset Splits Yes We randomly select 10,000 samples and another 10,000 samples as the validation and testing set, respectively. The remaining samples are all included in the training set.
Hardware Specification Yes Machine 1 for Mat Expert with Llama-2 8B and Llama-3 8B: Hardware: NVIDIA A5000 GPU (24 GB on-board memory) GPU x 4, AMD Ryzen Threadripper PRO 3975WX CPU, 256 GB RAM Machine 2 for Mat Expert with Llama-2 70B and Llama-3 70B: Hardware: NVIDIA L40S GPU (48 GB on-board memory) x 4, INTEL(R) XEON(R) GOLD 6526Y x 2, 512 GB RAM
Software Dependencies Yes Software: Ubuntu 22.04, Python 3.11, Py Torch 2.3.1+cu121, CUDA 12.4
Experiment Setup Yes Table 4: Hyperparameters for training T5-based encoders in the retrieval stage. Hyperparameter Value Description Model Architecture t5-base ... Learning Rate 1e-4 ... Number of epochs 100 ... Temperature 0.1 ... Batch Size 32 ... Gradient Accumulation Steps 8. Table 5: Hyperparameters for finetuning Llama family in the transition and generation stage. Hyperparameter Llama-3 8B Llama-3 70B Model Architecture Meta-Llama-3-8B-Instruct Meta-Llama-3-70B-Instruct Learning Rate 1e 4 1e 4 Batch Size 4 4 Batch Size per Device 1 1 Number of Epochs 70 10 Optimizer Adam W Adam W Sequence Length 2048 1024 Lo RA Rank 8 8 Lo RA Alpha 16 16 Lo RA Dropout 0.1 0.1 Warmup Ratio 0.1 0.1 Gradient Accumulation Steps 8 8