Multi-Modal Foundation · mist-models
MIST 1.8B · H298
A 1.8-billion parameter transformer predicting molecular enthalpy at 298.15 K from Kekulized SMILES strings.
What MIST 1.8B · H298 is
MIST 1.8B · H298 is a molecular property prediction model developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor. It adapts the MIST (Molecular Insight SMILES Transformers) 1.8-billion parameter foundation encoder to forecast the thermodynamic enthalpy of molecules at 298.15 K (H298).
Enthalpy reflects the total heat content of a chemical system and serves as a fundamental variable in predicting chemical reaction behaviour, thermochemical balances, and phase transitions. By mapping molecular SMILES representations directly to enthalpy predictions, this model provides rapid property estimation without requiring expensive quantum mechanical calculations.
The model builds upon a pre-trained encoder initially self-supervised on Enamine REAL Space chemical structures, which is subsequently fine-tuned using supervised regression on quantum chemical targets from the QM9 benchmark within MoleculeNet.
How it works
The base architecture is a RoBERTa-PreLayerNorm encoder consisting of 24 layers, a hidden dimension of 2,304, an intermediate size of 9,216, 18 attention heads, and a maximum context length of 2,048 tokens. During pre-training, it was trained on Kekulized SMILES from the Enamine REAL Space dataset using a Masked Language Modelling (MLM) objective with cross-entropy loss, optimised using FusedLAMB.
For downstream enthalpy prediction, the fine-tuned model couples the pre-trained encoder to a task network composed of a two-layer multi-layer perceptron (MLP) with Gaussian Error Linear Units (GELU) activations and dropout. Hidden states are pooled using the representation of the sequence's first token. The task network is fine-tuned on the QM9 dataset using mean squared error (MSE) loss and the AdamW optimiser to predict H298 values in Hartree.
Key facts
- Base architecture
- RoBERTa-PreLayerNorm with a 2-layer MLP head
- Layers
- 24
- Hidden size
- 2304
- Attention heads
- 18
- Maximum sequence length
- 2048 tokens
- Pre-training objective
- Masked Language Modelling (MLM)
- Pre-training dataset
- Enamine REAL Space
- Fine-tuning dataset
- QM9 Enthalpy at 298.15K (MoleculeNet)
- Fine-tuning loss
- Mean Squared Error (MSE)
- Input format
- Kekulized SMILES strings
- Output format
- Enthalpy at 298.15 K (in Hartree)
What you can use it for
- Thermodynamic Enthalpy Estimation
- Predict standard enthalpy at 298.15 K (H298) in Hartree directly from input molecular SMILES strings.
- High-Throughput Thermochemical Screening
- Evaluate large candidate libraries of small molecules to filter candidates based on thermodynamic stability and heat content.
- Molecular Representation Extraction
- Extract mean-pooled or initial-token hidden representations of molecular structures for downstream statistical analyses.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
MIST 1.8B · H298
Strengths and limits
Good at
- Pre-trained on Enamine REAL Space, capturing extensive structural patterns from synthetically accessible chemical space.
- Combines a large 1.8-billion parameter RoBERTa-PreLayerNorm encoder with a lightweight GELU-activated MLP head specialised for regression.
- Accommodates molecular SMILES sequences up to 2,048 tokens in length.
Watch out for
- Evaluated specifically using Mean Absolute Error on a random 80/10/10 split of the QM9 dataset, restricting validated accuracy to small organic molecules characteristic of QM9.
- Expects input molecules formatted as Kekulized SMILES strings and relies on custom tokenization requiring a Rust-based tokenizer.
- Requires substantial GPU resources, needing 24 GB of GPU memory for serving and inference.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · See model card
The See model card licence carries terms worth reading before you depend on it.
Questions
Can I use MIST 1.8B · H298 commercially?
Check the licence before commercial use. The See model card licence carries terms worth reading before you depend on it.
How do I run MIST 1.8B · H298 without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 16GB, typically in 12s, and a run costs about 70 credits. Visitors without an account get 1,000 credits a month.
What unit does the model use for enthalpy predictions?
The model outputs enthalpy at 298.15 K (H298) in Hartree.
What molecular input representation is required?
The model takes SMILES strings as inputs, which should be Kekulized prior to inference.
What tokenizer is used with this model?
It uses the SMIRK tokenizer (SmirkTokenizerFast), which requires Rust to be installed in the computational environment.
Which dataset was used to fine-tune the model?
The model was fine-tuned on the QM9 Enthalpy at 298.15K dataset from the MoleculeNet benchmark.
How to cite it
Wadell, Alexius and Bhutani, Anoushka and Azumah, Victor and Ellis-Mohr, Austin R. and Kelly, Celia and Zhao, Hancheng and Nayak, Anuj K. and Hegazy, Kareem and Brace, Alexander and Lin, Hongyi and Emani, Murali and Vishwanath, Venkatram and Gering, Kevin and Alkan, Melisa and Gibbs, Tom and Wells, Jack and Varshney, Lav R. and Ramsundar, Bharath and Duraisamy, Karthik and Mahoney, Michael W. and Ramanathan, Arvind and Viswanathan, Venkatasubramanian. Foundation Models for Discovery and Exploration in Chemical Space. 2025. arXiv:2510.18900 [physics]. https://doi.org/10.48550/arXiv.2510.18900
Written from MIST 1.8B · H298's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.