SuperconGPT¶

Three-tab superconductor discovery platform: (1) Inverse design — generate crystal structures for a target Tc using AtomGPT, optionally relax with ALIGNN-FF and predict properties, (2) Database search — query SuperconDB (1058 3D + 161 2D materials + hydrides) by elements with α²F(ω) Eliashberg function plots, (3) Tc predictor — predict superconducting Tc and 7 other properties from any POSCAR using ALIGNN.

Open App

Overview¶

SuperconGPT combines generative AI (AtomGPT) for inverse design of superconductors with a curated DFT-based BCS superconductor database and ALIGNN property prediction. The database includes Tc, electron-phonon coupling constant λ, logarithmic phonon frequency ω_log, α²F(ω) Eliashberg functions, and stability information for ambient-pressure 3D and 2D superconductors.

Data Source

SuperconDB 3D (1058 entries) + SuperconDB 2D (161 entries) + hydrides from jarvis.db.figshare. AtomGPT (SuperconGPT LoRA adapter) for inverse crystal generation. ALIGNN (jv_supercon_tc_alignn) for Tc prediction.

Endpoints¶

`POST /supercon/generate` — Inverse design: generate structure for target Tc¶

Generate a crystal structure for a given chemical formula and target superconducting Tc using AtomGPT. Optionally relax with ALIGNN-FF and run ALIGNN property predictions.

curl -X POST "https://atomgpt.org/supercon/generate" \
  -H "Authorization: Bearer sk-XYZ" \
  -H "Content-Type: application/json" \
  -H "accept: application/json" \
  -d '{
    "formula": "MgB2",
    "tc": 39.0,
    "max_new_tokens": 512,
    "temperature": 0.0,
    "top_k": 50,
    "top_p": 1.0,
    "do_sample": false,
    "relax": false,
    "run_alignn": true
  }'

Field	Type	Default	Description
`formula`	string	required	Chemical formula (e.g. `"MgB2"`, `"LaH10"`, `"H3S"`)
`tc`	float	required	Target superconducting Tc in Kelvin
`max_new_tokens`	int	512	Max generation tokens (64–2048)
`temperature`	float	0.0	Sampling temperature (0 = deterministic)
`top_k`	int	50	Top-K sampling (1–500)
`top_p`	float	1.0	Nucleus sampling threshold
`do_sample`	bool	false	Enable stochastic sampling
`relax`	bool	false	Relax generated structure with ALIGNN-FF
`run_alignn`	bool	false	Run ALIGNN property predictions on result

Response includes poscar, formula_generated, spacegroup, num_atoms, density, lattice, generation_time_s, xyz (for 3D viewer), and optionally alignn_predictions (formation_energy, bandgap, bulk_modulus, shear_modulus, supercon_tc_predicted).

`GET /supercon/generate` — Lightweight inverse design (GET)¶

Same as POST but via query parameters:

curl "https://atomgpt.org/supercon/generate?formula=MgB2&tc=39&relax=false&max_new_tokens=512" \
  -H "Authorization: Bearer sk-XYZ" \
  -H "accept: application/json"

`GET /supercon/search` — Search SuperconDB by elements¶

Search the DFT-based BCS superconductor database by element combination. Returns entries containing ALL specified elements, sorted by Tc descending.

curl "https://atomgpt.org/supercon/search?elements=Mg,B" \
  -H "Authorization: Bearer sk-XYZ" \
  -H "accept: application/json"

curl "https://atomgpt.org/supercon/search?elements=Nb" \
  -H "Authorization: Bearer sk-XYZ" \
  -H "accept: application/json"

curl "https://atomgpt.org/supercon/search?elements=La,H" \
  -H "Authorization: Bearer sk-XYZ" \
  -H "accept: application/json"

Param	Description
`elements`	Comma-separated elements (e.g. `"Mg,B"`, `"Nb"`, `"La,H"`)

Response includes count, elements, and entries array with: jid, formula, tc (K), lambda (electron-phonon coupling), wlog (K), pressure, stability, a2F_x, a2F_y (Eliashberg function data for plotting).

`POST /supercon/predict_tc` — Predict Tc from POSCAR¶

Predict superconducting Tc and 7 other properties from any crystal structure using ALIGNN pre-trained models. Max 50 atoms.

curl -X POST "https://atomgpt.org/supercon/predict_tc" \
  -H "Authorization: Bearer sk-XYZ" \
  -H "Content-Type: application/json" \
  -H "accept: application/json" \
  -d '{
    "poscar": "MgB2\n1.0\n1.537 -2.662 0.0\n1.537 2.662 0.0\n0.0 0.0 3.515\nMg B\n1 2\nCartesian\n0.0 0.0 0.0\n1.537 -0.887 1.757\n1.537 0.887 1.757"
  }'

Response includes formula, num_atoms, spacegroup, and properties:

Property	Unit	Description
`supercon_tc`	K	Predicted superconducting Tc
`formation_energy`	eV/atom	Formation energy
`total_energy`	eV/atom	Total energy
`bandgap_optb88`	eV	Band gap (OptB88vdW)
`bandgap_mbj`	eV	Band gap (mBJ)
`bulk_modulus`	GPa	Bulk modulus (Voigt)
`shear_modulus`	GPa	Shear modulus (Voigt)
`piezoelectric`	—	Max piezoelectric constant

Python Examples¶

Inverse designDatabase searchPredict Tc

import requests

response = requests.post(
    "https://atomgpt.org/supercon/generate",
    headers={
        "Authorization": "Bearer sk-XYZ",
        "accept": "application/json",
        "Content-Type": "application/json",
    },
    json={
        "formula": "MgB2",
        "tc": 39.0,
        "relax": True,
        "run_alignn": True,
    },
)
data = response.json()
if data["success"]:
    print(f"Generated: {data['formula_generated']} ({data['spacegroup']})")
    print(f"Atoms: {data['num_atoms']}, Time: {data['generation_time_s']}s")
    if "alignn_predictions" in data:
        ap = data["alignn_predictions"]
        print(f"Predicted Tc: {ap.get('supercon_tc_predicted', '?')} K")
    print(f"POSCAR:\n{data['poscar'][:300]}")

import requests

response = requests.get(
    "https://atomgpt.org/supercon/search",
    headers={
        "Authorization": "Bearer sk-XYZ",
        "accept": "application/json",
    },
    params={"elements": "Nb"},
)
data = response.json()
print(f"Found {data['count']} Nb superconductors")
for e in data["entries"][:10]:
    print(f"  {e['jid']:12s} {e['formula']:10s} Tc={e['tc']} K  λ={e['lambda']}")

import requests

MGB2_POSCAR = """MgB2
1.0
1.537 -2.662 0.0
1.537 2.662 0.0
0.0 0.0 3.515
Mg B
1 2
Cartesian
0.0 0.0 0.0
1.537 -0.887 1.757
1.537 0.887 1.757"""

response = requests.post(
    "https://atomgpt.org/supercon/predict_tc",
    headers={
        "Authorization": "Bearer sk-XYZ",
        "accept": "application/json",
        "Content-Type": "application/json",
    },
    json={"poscar": MGB2_POSCAR},
)
data = response.json()
if data["success"]:
    p = data["properties"]
    print(f"{data['formula']} ({data['spacegroup']})")
    print(f"  Predicted Tc: {p['supercon_tc']:.2f} K")
    print(f"  Formation energy: {p['formation_energy']:.3f} eV/atom")
    print(f"  Band gap: {p['bandgap_optb88']:.3f} eV")

AGAPI Agent [WIP]¶

from agapi.agents import AGAPIAgent
import os

agent = AGAPIAgent(api_key=os.environ.get("AGAPI_KEY"))

# Inverse design
response = agent.query_sync("Generate a crystal structure for MgB2 with Tc=39K using SuperconGPT")
print(response)

# Database search
response = agent.query_sync("Search for Nb-based superconductors in SuperconDB")
print(response)

# Predict Tc
response = agent.query_sync("Predict the superconducting Tc for JVASP-1002")
print(response)

References¶

K. Choudhary, K. Garrity, npj Comp. Mat. 8, 244 (2022) — Designing High-Tc Superconductors with BCS-Inspired Screening DOI
K. Choudhary, J. Phys. Chem. Lett. 15, 6909 (2024) — AtomGPT DOI
D. Wines, T. Xie, K. Choudhary, J. Phys. Chem. Lett. 14, 6630 (2023) — Inverse Design of Superconductors DOI
D. Wines, K. Choudhary et al., Nano Lett. 23, 969 (2023) — 2D Superconductors DOI
D. Wines, K. Choudhary, Materials Futures 3, 025602 (2024) — High-Pressure Hydride Superconductors DOI
Charles Rhys Campbell, Aldo H. Romero, and Kamal Choudhary, AtomBench: A Benchmarking Framework for Generative Crystal Reconstruction Models in Conventional Superconductors DOI
atomgptlab/atomgpt
atomgptlab/alignn
atomgptlab/atombench_inverse