Multi-Modal Foundation · mist-models
MIST mixtures
Predicts thermodynamic excess properties and mixing behaviour of binary chemical mixtures from molecular SMILES.
What MIST mixtures is
MIST mixtures is a molecular property prediction model developed by the Electrochemical Energy Group at the University of Michigan, Ann Arbor. It is part of the Molecular Insight SMILES Transformers (MIST) family, built on the MIST-28M encoder architecture.
The model addresses the challenge of predicting the thermodynamic behaviour of binary chemical mixtures across various compositions and temperatures. In thermodynamics, excess properties represent the difference between the actual thermodynamic properties of a mixture and those of an ideal solution. MIST mixtures calculates density, molar volume, and molar enthalpy, decomposing them into linear mixing components and non-ideal excess terms.
How it works
The underlying encoder is a RoBERTa-PreLayerNorm architecture comprising 8 layers, 8 attention heads, a hidden size of 512, an intermediate size of 2048, and a maximum sequence length of 2048 tokens. It was pretrained on SMILES strings from the Enamine REAL Space dataset using a masked language modelling objective with cross-entropy loss and the FusedLAMB optimiser.
For mixture property prediction, single-molecule embeddings generated by the encoder are combined using a permutation-equivariant fusion function. A two-layer multilayer perceptron (MLP) task network predicts excess property values across fixed mole ratio control points, which are then used to fit an interpolating polynomial that explicitly constrains excess properties to zero at the pure endpoints. A second MLP estimates pure compound properties to calculate linear mixing contributions. The task network was fine-tuned with AdamW using a multi-channel mean squared error loss over absolute and excess properties.
Key facts
- Architecture
- RoBERTa-PreLayerNorm encoder (8 layers, 8 heads, hidden size 512) with a two-layer MLP task network
- Parameter count
- 28 million (encoder)
- Pretraining data
- Enamine REAL Space SMILES dataset
- Fine-tuning data
- Excess Property Dataset (888,045 observations across 715 molecules and 1,519 binary mixtures)
- Context length
- 2048 tokens
What you can use it for
- Binary mixture excess property prediction
- Predict excess density, molar volume, and molar enthalpy for two-component chemical solutions at defined mole fractions and temperatures in Kelvin.
- Linear and absolute property estimation
- Calculate both pure-component linear mixing baselines and total absolute mixture properties for liquid formulations.
- Chemical sequence embedding
- Generate mean-pooled representation vectors from chemical sequence inputs for downstream feature-extraction tasks.
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 mixtures
Strengths and limits
Good at
- Mathematically guarantees that excess properties equal zero at pure substance endpoints through polynomial interpolation.
- Preserves permutation invariance regardless of the ordering of the two chemical components in the input.
- Predicts excess, linear mixing, and overall absolute values simultaneously in a single model forward pass.
Watch out for
- Restricted to binary chemical mixtures and cannot model systems containing three or more chemical components.
- Fine-tuning coverage is limited to literature data spanning 715 distinct molecules and 1,519 unique mixture pairs.
- Requires the SMIRK tokenizer, which relies on a local Rust installation.
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 mixtures 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 mixtures 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 CPU only, typically in 12s, and a run costs about 28 credits. Visitors without an account get 1,000 credits a month.
What inputs does the model require for property prediction?
The model expects a list of two SMILES strings representing the binary components, their respective mole fractions summing to one, and the temperature in Kelvin.
What thermodynamic properties are output by MIST mixtures?
It outputs predictions for density in grams per cubic centimetre, molar volume in cubic centimetres per mole, and molar enthalpy in joules per mole, broken down into absolute, linear, and excess terms.
What base checkpoint was used for fine-tuning?
The model was fine-tuned from the pretrained encoder checkpoint mist-28M-ti624ev1.
How are mixture components fused during prediction?
Individual molecular embeddings from the MIST encoder are combined using a permutation-equivariant fusion function before being processed by the property multilayer perceptron.
How to cite it
Wadell, A., Bhutani, A., Azumah, V., Ellis-Mohr, A. R., Kelly, C., Zhao, H., Nayak, A. K., Hegazy, K., Brace, A., Lin, H., Emani, M., Vishwanath, V., Gering, K., Alkan, M., Gibbs, T., Wells, J., Varshney, L. R., Ramsundar, B., Duraisamy, K., Mahoney, M. W., Ramanathan, A., and Viswanathan, V. Foundation Models for Discovery and Exploration in Chemical Space. arXiv:2510.18900 (2025). https://doi.org/10.48550/arXiv.2510.18900
Written from MIST mixtures'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.