MaxKannapinn/chicken_dataset_comsol

Chicken Breast Cooking Simulation Dataset (chicken_dataset.nc)

This repository hosts a comprehensive dataset of 566 transient finite-element simulations modeling the coupled heat and moisture transfer during convective roasting of chicken breast meat. Originally solved using high-fidelity numerical formulations in COMSOL Multiphysics, the dataset is structured as a CF-compliant, hierarchical NetCDF-4 / HDF5 store (xarray.DataTree).

---

1. Physical Context & Parametric Variation

What Changes from Simulation to Simulation?

The geometric domain, physical material properties, and initial internal states are standardized across all runs. The only parameter that changes from simulation to simulation is the oven temperature profile over time ($T_{\text{oven}}(t)$).

Each experiment represents a unique thermal processing trajectory:

• Different heating regimes: multi-step ramps, holds, harmonic oscillations (amplitude/frequency sweeps), and dynamic step-changes.
• Temperatures vary across realistic culinary and industrial thermal processing ranges ($\approx 279\text{ K}$ up to $473\text{ K}$, or $6^\circ\text{C}$ to $200^\circ\text{C}$).
• Each run spans approximately $1395\text{ seconds}$ ($\approx 23.25\text{ minutes}$) sampled at $5\text{ s}$ intervals ($280$ to $281$ discrete timesteps).

This structure makes the dataset ideal for:

• Reduced-Order Modeling (ROM) and dynamic system identification (DynROM, DeepONet, FNO).
• Physics-Informed Neural Networks (PINNs) and Neural ODEs.
• Multivariate time-series forecasting under exogenous control inputs.

---

2. NetCDF-4 Hierarchical Data Structure

The dataset (chicken_dataset_v73.nc) is organized hierarchically where each simulation is an independent root-level group /{exp_name} (e.g., /272-a3nl473h300t1400, /1037-m60amp10t1400schroeder).

chicken_dataset_v73.nc
└── /{exp_name}                                [Root experiment group]
    │   Dimensions:   (time: 280)
    │   Coordinates:  time                     [s, range: 0 to 1395]
    │   Attributes:   run_name, session, type, {exp_name}_oven_temperature.csv
    │
    ├── Data Variables (1D Global & Volume-Averaged Signals):
    │     ├── T_oven                           [K] Global oven temperature profile
    │     ├── Temperature_volavg_K             [K] Volume-averaged domain temperature
    │     └── massConcentrationWater_volavg_1_ [-] Volume-averaged water mass fraction
    │
    ├── /all_probes_measurements               [3-Probe Sensor Mesh Group]
    │     Dimensions:   (time: 280, node: 3, spatial: 3)
    │     Coordinates:
    │       ├── coordinates (node, spatial)    [(0, 0.010, 0), (0, 0.019, 0), (0, 0.020, 0)]
    │       └── spatial                        ['x', 'y', 'z'] (Cartesian meters)
    │     Inherited:    time                   [s]
    │     Variables (32): 3-point stacked continuum state variables
    │
    └── /single_probe_measurements             [Surface Boundary Mesh Group]
          Dimensions:   (time: 280, node: 1, spatial: 3)
          Coordinates:
            ├── coordinates (node, spatial)    [(0, 0.020, 0)]
            └── spatial                        ['x', 'y', 'z'] (Cartesian meters)
          Inherited:    time                   [s]
          Variables (6): Surface exchange flux & boundary condition quantities

---

3. Discretization & Variable Breakdown

A. Root Level: Global & Volume-Averaged Signals

These variables have only dimension (time,) with no node dimension, indicating they describe the whole domain or external boundary environment rather than a local probe coordinate:

Variable Name	Dimensions	Units	Description
T_oven	(time: 280)	$\text{K}$	Instantaneous oven heating schedule / control boundary temperature.
Temperature_volavg_K	(time: 280)	$\text{K}$	Volume-averaged temperature of the entire chicken breast $\langle T \rangle_V$.
massConcentrationWater_volavg_1_	(time: 280)	$-$	Volume-averaged moisture mass concentration $\langle C_w \rangle_V \in [0, 1]$.

B. Group /all_probes_measurements (3 Physical Probes)

Evaluated across three Cartesian probe locations:

1. $\mathbf{x}_1 = (0.000, 0.010, 0.000)\text{ m}$: Core / geometric center
2. $\mathbf{x}_2 = (0.000, 0.019, 0.000)\text{ m}$: Sub-surface layer
3. $\mathbf{x}_3 = (0.000, 0.020, 0.000)\text{ m}$: Outer surface boundary

Each variable is shaped as (time: 280, node: 3). The probe coordinate array is available under coordinates (node: 3, spatial: 3) with spatial = ['x', 'y', 'z'].

Key fields include:

• Temperature & Phase States: Temperature-K, C_eq.
• Constituent Volume Fractions: Volumefractionwater-1, Volumefractionfat-1, Volumefractionprotein-1, Volumefractionash-1.
• Mass Concentrations: concentraitonwater-mass--1, concentrationfat-mass--1, concentrationprotein-mass--1, concentrationash-mass--1, massconcentrationwater-1.
• Thermophysical Properties: Densitychickenbreast-kg_m^3, heatcapchickenmeat-J_-kg*K, k_para-W_-m*K, k_perp-W_-m*K, conduct.water-W_-m*K, heatcapwater-J_-kg*K.
• Fluid & Vapor Kinematics: umag-m_s, u_x_rab-m_s, u_y_rab-m_s, u_z_rab-m_s, storagemodulus-N_m^2.

(Note: Trailing --Point:... suffixes from raw tables have been stripped for clean array slicing).

C. Group /single_probe_measurements (Surface Boundary Probe)

Evaluated strictly at the surface probe $\mathbf{x}_3 = (0.000, 0.020, 0.000)\text{ m}$ (outer skin). Each variable is shaped as (time: 280, node: 1).

Variable Name	Dimensions	Units	Physical Interpretation
Step5betaskin-Point:-0-0.02-0-	(time, node)	$\text{m/s}$	Skin mass transfer coefficient step function.
beta_tot-m_s--Point:-0-0.02-0-	(time, node)	$\text{m/s}$	Total surface mass convective exchange coefficient.
evaporationmassfluxair-kg_-m^2*s---Point:-0-0.02-0-	(time, node)	$\text{kg}/(\text{m}^2\cdot\text{s})$	Mass flux of water vapor lost to the oven atmosphere.
wateractivity-1--Point:-0-0.02-0-	(time, node)	$-$	Surface thermodynamic water activity $a_w \in [0, 1]$.
watervapourconcentrationfilm-kg_m^3--Point:-0-0.02-0-	(time, node)	$\text{kg/m}^3$	Water vapor concentration in the boundary air film.
watervapoursaturationpressure-Pa--Point:-0-0.02-0-	(time, node)	$\text{Pa}$	Saturated water vapor pressure at surface temperature.

---

4. How to Use the Dataset in Python

NetCDF-4 files on Hugging Face are stored as binary Git LFS files. You can download and stream them directly into xarray.DataTree using huggingface_hub and h5netcdf.

Installation

pip install huggingface_hub xarray h5netcdf numpy matplotlib

Loading and Inspecting an Experiment

from huggingface_hub import hf_hub_download
import xarray as xr

# 1. Download NetCDF file from Hugging Face Hub cache
file_path = hf_hub_download(
    repo_id="your-username/chicken-cooking-dataset",
    filename="chicken_dataset_v73.nc",
    repo_type="dataset",
)

# 2. Open an individual simulation experiment
exp_name = "272-a3nl473h300t1400"

# Open root experiment (oven temperature and volume averages)
ds_root = xr.open_dataset(file_path, group=f"/{exp_name}", engine="h5netcdf")
print("Oven schedule:", ds_root["T_oven"])
print("Average core temperature:", ds_root["Temperature_volavg_K"])

# Open 3-probe measurements
ds_probes = xr.open_dataset(file_path, group=f"/{exp_name}/all_probes_measurements", engine="h5netcdf")
print("Core vs Surface Temperature shape:", ds_probes["Temperature-K"].shape)  # (280, 3)

# Open surface exchange fluxes
ds_surf = xr.open_dataset(file_path, group=f"/{exp_name}/single_probe_measurements", engine="h5netcdf")
print("Evaporative mass flux:", ds_surf["evaporationmassfluxair-kg_-m^2*s---Point:-0-0.02-0-"])

Plotting Core vs. Surface Dynamics

import matplotlib.pyplot as plt

time = ds_root["time"].values
t_oven = ds_root["T_oven"].values
t_core = ds_probes["Temperature-K"].sel(node=0).values     # Point (0, 0.010, 0)
t_surface = ds_probes["Temperature-K"].sel(node=2).values  # Point (0, 0.020, 0)

plt.figure(figsize=(9, 5))
plt.plot(time, t_oven, 'k--', label="Oven Temperature (Exogenous Input)")
plt.plot(time, t_core, 'b-', label="Chicken Core (Node 0)")
plt.plot(time, t_surface, 'r-', label="Chicken Surface (Node 2)")
plt.xlabel("Time [s]")
plt.ylabel("Temperature [K]")
plt.title(f"Thermal Response Curve: {exp_name}")
plt.grid(True, alpha=0.3)
plt.legend()
plt.tight_layout()
plt.show()

---

5. Metadata & Embedded Files

Each experiment group embeds its full numerical oven setpoint table directly in the group attributes under {exp_name}_oven_temperature.csv. You can extract it directly from the group metadata:

csv_text = ds_root.attrs[f"{exp_name}_oven_temperature.csv"]
print(csv_text[:200])

---

6. Citation & Academic Works

This dataset was created and published by Dr.-Ing. Maximilian Kannapinn (Technische Universität Darmstadt).

If you use this dataset or its numerical benchmarks in your research, please cite the corresponding publications:

@article{kannapinn2022physics,
  title={Physics-based digital twins for autonomous thermal food processing: Efficient, non-intrusive reduced-order modeling},
  author={Kannapinn, Maximilian and Pham, Minh Khang and Sch{\"a}fer, Michael},
  journal={Innovative Food Science \& Emerging Technologies},
  volume={81},
  pages={103143},
  year={2022},
  publisher={Elsevier},
  doi={10.1016/j.ifset.2022.103143}
}

@phdthesis{kannapinn2023digital,
  title={Digital Twins for Autonomous Thermal Food Processing},
  author={Kannapinn, Maximilian},
  year={2023},
  school={Technische Universit{\"a}t Darmstadt},
  address={Darmstadt, Germany},
  doi={10.26083/tuprints-00024386}
}

@article{kannapinn2025digital,
  title={Digital twins for autonomous thermal food processing: A model predictive control study with reduced-order models of augmented neural ordinary differential equation type},
  author={Kannapinn, Maximilian and Dorer, Denis and Sch{\"a}fer, Michael and Weeger, Oliver},
  journal={Journal of Food Engineering},
  year={2025},
  publisher={Elsevier},
  doi={10.1016/j.jfoodeng.2025.112918}
}