SoftnessML API

Machine learning framework for predicting particle rearrangements in supercooled liquids and glasses from local structure — served as a FastAPI microservice.

---

Overview

SoftnessML API predicts the "softness" of particles in disordered systems — a machine-learning-derived scalar field that captures local structural environments and predicts which particles are likely to rearrange.

Why this matters: In supercooled liquids and glasses, dynamics are highly heterogeneous. Some particles rearrange frequently while others stay static. Softness bridges the gap between static structure and dynamics, enabling prediction of rearrangements from a single snapshot.

What This Project Provides

• Calculation of local structural descriptors (radial, angular, and bond-orientational order parameters via spherical harmonics) from MD trajectories in GSD format
• Trained Support Vector Machine (SVM) model for softness prediction
• FastAPI REST API for integration into simulation pipelines
• Numba-accelerated computation for performance
• Docker deployment for reproducible environments

The implementation follows Schoenholz et al. (2016).

---

Results

Particle Trajectories Colored by Softness

Particles colored by their predicted softness field — warmer colors indicate higher softness (more likely to rearrange), cooler colors indicate lower softness (more stable).

3D Lennard-Jones supercooled liquid at T=0.63. Color scale: blue (low softness, stable) → red (high softness, rearrangement-prone).

Softness Distribution: Rearranging vs. Non-Rearranging Particles

The distribution of softness for particles that rearrange — $P(S|R)$ — is shifted to higher values compared to particles that remain static demonstrating that softness is a strong predictor of rearrangement propensity.

The clear separation between the two distributions confirms that the SVM-learned softness field successfully captures structural features predictive of dynamics.

---

Scientific Background

What is Softness?

In supercooled liquids and glasses, particle dynamics are highly heterogeneous. Softness is a machine-learning-derived scalar field that captures the local structural environment of each particle and predicts its likelihood to undergo a rearrangement.

The concept was introduced by Schoenholz et al. (2016) and has become a widely used tool for linking structure to dynamics in amorphous materials.

Applications

• Fundamental Physics: Glass transition and dynamic heterogeneity
• Materials Science: Stable amorphous materials, mechanical properties
• Industrial: Predicting aging, failure, and deformation in glassy systems

---

Mathematical Framework

Local Structure Descriptors

The local environment of each particle is characterized by three types of descriptors:

Radial Structure Functions

Capture local density variations at different distances:

$$ G_\mu(i) = \sum_{j \in \text{neighbors}} e^{-(r_{ij} - \mu)^2 / L^2} $$

where $\mu$ are radial distance parameters and $L$ is a characteristic length scale.

Angular Structure Functions

Capture three-body correlations:

$$ \Psi_{\xi,\lambda,\zeta}(i) = \sum_{j,k \in \text{neighbors}} e^{-\xi^2(r_{ij}^2 + r_{ik}^2 + r_{jk}^2)} (1 + \lambda \cos\theta_{jik})^\zeta $$

where $\theta_{jik}$ is the angle at particle $i$ formed by neighbors $j$ and $k$, and $\xi$, $\lambda$, $\zeta$ control radial decay and angular sensitivity.

Bond-Orientational Order Parameters (Steinhardt Parameters)

Quantify local rotational symmetry using spherical harmonics. For a central particle $i$, neighbors within annular shells (inner radii $r_\text{inner}$, width $\Delta r = 0.5$) are considered.

For each shell and each even angular momentum $l = 2, 4, 6, 8, 10, 12, 14$:

$$ q_{lm}(i) = \frac{1}{N_b(i)} \sum_{j \in \text{shell}} Y_{lm}(\theta_{ij}, \phi_{ij}) $$

$$ q_l(i) = \sqrt{\frac{4\pi}{2l+1} \sum_{m=-l}^{l} |q_{lm}(i)|^2} $$

where $N_b(i)$ is the number of neighbors in the shell, $Y_{lm}(\theta, \phi)$ are spherical harmonics, and $(\theta_{ij}, \phi_{ij})$ are the polar and azimuthal angles of the vector from particle $i$ to neighbor $j$.

Softness Prediction (SVM)

Softness is computed as a linear combination of structural descriptors using a trained SVM:

$$ S_i = \mathbf{w} \cdot \mathbf{x}_i + b $$

where $\mathbf{x}$ is the concatenated feature vector, $\mathbf{w}$ is the learned weight vector, and $b$ is the bias term.

Rearrangement Detection

The SVM is trained to separate particles that undergo significant motion using the hop parameter $p_\text{hop}$ Candelier et al.:

$$ p_\text{hop}(i,t) = \sqrt{\langle(\vec{r}_i(t) - \langle\vec{r}_i\rangle_{w_2})^2\rangle_{w_1} \langle(\vec{r}_i(t) - \langle\vec{r}_i\rangle_{w_1})^2\rangle_{w_2}} $$

where time windows are $w_1 = [t-5, t]$ and $w_2 = [t, t+5]$ (in Lennard-Jones time units).

---

Quick Start

Prerequisites

• Python 3.9+
• Docker (recommended for deployment)

Option 1: Docker (Recommended)

# Clone
git clone https://github.com/tomidiy/SoftnessML_api.git
cd SoftnessML_api

# Build and run
docker build -t softness-predictor .
docker run -d -p 8000:8000 \
    -v $(pwd)/data:/app/data \
    --name softness-predictor-container \
    softness-predictor

# Health check
curl http://localhost:8000/health
# → {"status": "healthy"}

# Predict softness
curl -X POST "http://localhost:8000/predict" \
    -H "Content-Type: application/json" \
    -d '{"temp": 0.7, "frame": 0, "gsd_file": "T0.7_idx0.gsd"}'

Option 2: Local Development

git clone https://github.com/tomidiy/SoftnessML_api.git
cd SoftnessML_api
pip install -r app/requirements.txt

# Start the API server
uvicorn app.main:app --host 0.0.0.0 --port 8000

# Test
curl http://localhost:8000/health

Required Data Files

Place the following in data/:

data/
├── Softness_train_data.pkl
├── Softness_train_data_radialSphericalHarmonics.pkl
├── phop_T<temp>.pkl
└── T<temp>_idx0.gsd

---

API Reference

| Endpoint | Method | Description | |

|

|

| | /health | GET | Returns service health status | | /predict | POST | Compute descriptors and predict softness |

POST /predict

Computes structural descriptors (radial, angular, and spherical-harmonics-based bond-orientational parameters) from the specified GSD file and frame, then predicts softness using the trained SVM.

Request:

{
    "temp": 0.7,
    "frame": 0,
    "gsd_file": "T0.7_idx0.gsd"
}

| Field | Type | Description | |

|

|

| | temp | float | Simulation temperature | | frame | int | Trajectory frame index | | gsd_file | string | GSD file name in data/ |

Response:

{
    "softness": [0.42, -1.13, 0.87, ...],
    "num_particles": 1000
}

---

Project Structure

SoftnessML_api/
├── app/
│   ├── main.py               # FastAPI application and endpoints
│   ├── model.py               # SVM model loading and prediction
│   ├── Structure.py           # Structural descriptor computation
│   └── requirements.txt       # Python dependencies
├── data/
│   ├── Softness_train_data.pkl
│   ├── Softness_train_data_radialSphericalHarmonics.pkl
│   ├── phop_T<temp>.pkl
│   └── T<temp>_idx0.gsd
├── Dockerfile
├── README.md
└── .github/workflows/ci.yml   # CI pipeline

---

Tech Stack

| Component | Technology | Purpose | |

|

|

| | API Framework | FastAPI | REST endpoint serving | | ML Model | scikit-learn (SVM) | Softness prediction | | Descriptors | NumPy + SciPy + Numba | Structure function computation | | Trajectory I/O | GSD | Read MD simulation files | | Deployment | Docker | Reproducible containerized service | | CI | GitHub Actions | Automated testing |

---

Related Projects

• Multi-Modal RAG over Scientific Papers — RAG system for querying scientific papers using text, figures, and tables.

---

References

1. Schoenholz, S. S., et al. (2016). A structural approach to relaxation in glassy liquids. Nature Physics, 12, 469–471.
2. Cubuk, E. D., et al. (2015). Identifying structural flow defects in disordered solids using machine-learning methods. Physical Review Letters, 114, 108001.
3. Steinhardt, P. J., Nelson, D. R., & Ronchetti, M. (1983). Bond-orientational order in liquids and glasses. Physical Review B, 28, 784.

---

License

MIT License — see LICENSE.