torchcor 0.3.2

High-Performance Large-Scale Cardic Electrophysiology Simulations on GPUs

TorchCor

TorchCor is a high-performance simulator for cardiac electrophysiology (CEP) using the finite element method (FEM) on general-purpose GPUs. Built on top of PyTorch, TorchCor delivers substantial computational acceleration for large-scale CEP simulations, with seamless integration into modern deep learning workflows and efficient handling of complex mesh geometries.

TorchCor offers:

• 🚀 Fast, scalable CEP simulations on large and complex heart meshes
• 🔗 Seamless integration with PyTorch and scientific machine learning workflows
• ⚙️ Support for a wide range of ionic models and conductivity heterogeneity
• 🔧 Fully customizable model parameters for flexible experimentation and prototyping
• 🎯 Accurate simulation of cardiac electrical activity for research and development
• 📈 Generation of precise local activation and repolarization time maps
• 🩺 Simulation of clinically relevant 12-lead ECG signals through phie recovery (under testing)

🫀 Simulation Previews

Below are simulation results showcasing the electrical activation patterns over time in the left atrium and bi-ventricle.

3D left atrium surface mesh

3D bi-ventricle volume mesh

⚡ Performance

TorchCor is optimized for high-throughput cardiac electrophysiology simulations on large-scale meshes. The benchmarks below demonstrate its ability to efficiently scale with mesh size and leverage GPU acceleration over traditional CPU-based solvers. Performance tests were conducted using an AMD Ryzen Threadripper 3990X 64-Core Processor and the following GPUs:

• NVIDIA Tesla V100
• NVIDIA GeForce RTX 3090
• NVIDIA RTX A6000
• NVIDIA A100 80GB PCIe
• NVIDIA H100 80GB HBM3

Execution time on cubic 3D volume meshes with increasing node counts.

Execution time on a bi-ventricle mesh (637,480 nodes) using various CPU cores and GPU devices.

Unlike traditional CPU-based solvers like PETSc, which rely heavily on MPI-based parallelism and incur communication overhead, TorchCor minimizes latency by exploiting GPU-local memory and massive parallelism. This leads to superior scaling on large meshes, where CPU frameworks struggle with inter-process communication and abstraction overheads, allowing a high-throughput, low-latency pipeline well-suited for time-sensitive cardiac simulations.

🚀 Quickstart Example

Here’s a concise example to run a simulation using the TenTusscher-Panfilov ionic model on a bi-ventricle mesh.

🧩 Mesh file:
You can download the bi-ventricle mesh from the following link:
Google Drive – Bi-ventricle Mesh

The inputs are:

• .pts file: coordinates of the vertices
• .elem file: connectivities of the mesh
• .lon file: fibre orientations
• One or more .vtx files: pacing locations

Running the following code will produce:

• a list of membrane potentials, each saved after every 1 ms of the simulation
• a local activation time map and a repolarisation time map

all saved in .pt file format readable by torch.load.

import torchcor as tc
from torchcor.simulator import Monodomain
from torchcor.ionic import TenTusscherPanfilov
from pathlib import Path

# Specify the GPU device for running the simulation
tc.set_device("cuda:1")
dtype = tc.float32
# The total simulation duration (ms)
simulation_time = 600
dt = 0.01

home_dir = Path.home()
mesh_dir = home_dir / "Data/ventricle/Case_1"
# Load in the ionic model. Here we use TenTussherPanfilov for the simulation on bi-ventricle
ionic_model = TenTusscherPanfilov(cell_type="ENDO", dt=dt, dtype=dtype)
# 1. Initialise the Mondomain model
simulator = Monodomain(ionic_model, T=simulation_time, dt=dt, dtype=dtype)
# 2. Load in the mesh files (.pts .elem .lon)
simulator.load_mesh(path=mesh_dir)
# 3. Specify the conductivity for each region
simulator.add_conductivity([34, 35], il=0.5272, it=0.2076, el=1.0732, et=0.4227)
simulator.add_conductivity([44, 45, 46], il=0.9074, it=0.3332, el=0.9074, et=0.3332)
# 4. Specify the locations where stimulation is applied
simulator.add_stimulus(mesh_dir / "LV_sf.vtx", start=0.0, duration=1.0, intensity=100)
simulator.add_stimulus(mesh_dir / "LV_pf.vtx", start=0.0, duration=1.0, intensity=100)
simulator.add_stimulus(mesh_dir / "LV_af.vtx", start=0.0, duration=1.0, intensity=100)
simulator.add_stimulus(mesh_dir / "RV_sf.vtx", start=5.0, duration=1.0, intensity=100)
simulator.add_stimulus(mesh_dir / "RV_mod.vtx", start=5.0, duration=1.0, intensity=100)

# 5. Start the simulation
simulator.solve(a_tol=1e-5,                  # absolute tolerance in CG
                r_tol=1e-5,                  # relative tolerance
                max_iter=100,                # maximum number of iterations for each CG calculation
                calculate_AT_RT=True,        # keep track of local activation time (LAT)
                linear_guess=True,
                snapshot_interval=1,         # save the solution after every 1 ms
                verbose=True,
                result_path="./biventricle") # the folder in which the results are saved

📦 Installation

pip install torchcor

Note: Requires PyTorch with CUDA support for GPU acceleration.

🐳 Docker Support

For a GPU-enabled Docker setup to run TorchCor without installing dependencies on your host system, please see the docker/ folder.

📖 Citation

If you use TorchCor in academic work, please consider citing:

@article{zhou2026torchcor,
  title={TorchCor: High-performance cardiac electrophysiology simulations with the finite element method on GPUs},
  author={Zhou, Bei and Balmus, Maximilian and Corrado, Cesare and Cicci, Ludovica and Qian, Shuang and Niederer, Steven A},
  journal={SoftwareX},
  volume={33},
  pages={102521},
  year={2026},
  publisher={Elsevier}
}

👩‍💻 Contributors

TorchCor is developed and maintained by Bei Zhou, Maximilian Balmus, Cesare Corrado, Shuang Qian, and Steven A. Niederer​ in the Cardiac Electro-Mechanics Research Group (CEMRG) at Imperial College London.

We welcome contributions from the community! Feel free to open issues or submit pull requests.