PEA-PGNN

Documentation · Examples · Public API · Model card · Contributing · Issues · Releases

Prior-anchored, structure-preserving neural prediction for time-dependent engineering responses.

PEA-PGNN is a research-oriented Python package for prior-anchored, structure-preserving prediction of time-dependent engineering responses. It turns three types of incomplete engineering knowledge into explicit computational roles:

1. quantitative empirical estimates become correctable parameter anchors;
2. alternative temporal laws form a context-conditioned convex mixture; and
3. non-negativity, monotonicity, and boundedness are inherited from the forward construction.

The first application bundled with the package is long-term concrete drying shrinkage prediction. The package contains the method and empirical-prior utilities, but does not contain the paper's database, trained weights, or submission files.

Status: 0.1.2 alpha. This repository is a clean, reusable implementation extracted from research code. It is not yet the exact reproduction archive for every experiment reported in the manuscript.

Highlights

• Prior-anchored learning: empirical estimates enter the model as correctable magnitude and timescale anchors rather than fixed answers.
• Structure-preserving evolution: convex mixtures of normalized temporal laws inherit non-negativity, monotonicity, and boundedness by construction.
• Two levels of use: work with the low-level PyTorch module or the fitted NumPy-style regressor interface.
• Concrete application: use B3-, GL2000-, and ACI209-inspired utilities to construct drying-shrinkage priors.
• Auditable behavior: check predicted trajectories and standard regression metrics with reusable evaluation utilities.
• Research-oriented documentation: explicit input contracts, method boundaries, reproducibility scope, tests, and a runnable synthetic example.

Installation

Requirements

• Python 3.9 or later;
• NumPy 1.23 or later;
• scikit-learn 1.2 or later; and
• PyTorch 2.0 or later.

The runtime dependencies are installed automatically. A GPU is optional; for a CUDA- or ROCm-specific PyTorch build, follow the official PyTorch installation selector before installing PEA-PGNN.

Install from PyPI

Install the released package with:

python -m pip install pea-pgnn

Install from GitHub

The latest development version can be installed directly from GitHub:

python -m pip install "git+https://github.com/hunter137/pea-pgnn.git"

Alternatively, clone the repository and install it locally:

git clone https://github.com/hunter137/pea-pgnn.git
cd pea-pgnn
python -m pip install -e .

Development install

For development and testing:

python -m pip install -e ".[dev]"
python -m pytest

Package naming

The distribution name contains a hyphen, while the Python import name uses an underscore:

import pea_pgnn

Quick start: concrete empirical anchors

Inputs may be scalars or broadcast-compatible NumPy arrays. Shrinkage is returned in microstrain.

from pea_pgnn.concrete import concrete_prior_anchors

priors = concrete_prior_anchors(
    loading_age=7.0,          # d
    relative_humidity=60.0,  # %
    volume_surface_ratio=50.0,  # mm
    water_content=180.0,     # kg/m^3
    compressive_strength=40.0,  # MPa
)

print(priors["magnitude"])
print(priors["timescale"])

concrete_prior_anchors returns magnitude, timescale, and the three component estimates b3_magnitude, gl2000_magnitude, and aci209_magnitude. Their formulation lineage, units, simplifications, and applicability limits are documented in docs/empirical-priors.md.

Quick start: structured temporal evolution

The four normalized candidate laws can also be used without training a neural network:

import numpy as np

from pea_pgnn import convex_time_evolution

time = np.array([0.0, 7.0, 28.0, 90.0, 365.0])
evolution = convex_time_evolution(
    time=time,
    timescale=80.0,
    alpha=0.5,
    weights=[0.25, 0.25, 0.25, 0.25],
)

prediction = 650.0 * evolution

Quick start: train a prior-anchored model

PriorAnchoredRegressor offers a compact NumPy-style interface. Context must contain only time-invariant condition descriptors; query time is passed separately so that mixture weights remain fixed along one condition trajectory.

from pea_pgnn import PriorAnchoredRegressor, TrainingConfig

regressor = PriorAnchoredRegressor(
    training_config=TrainingConfig(epochs=300, patience=40, seed=42)
)

regressor.fit(
    context=X_train,
    time=t_train,
    target=y_train,
    magnitude_prior=A_prior_train,
    timescale_prior=tau_prior_train,
)

y_pred = regressor.predict(
    context=X_test,
    time=t_test,
    magnitude_prior=A_prior_test,
    timescale_prior=tau_prior_test,
)

A complete runnable synthetic example is provided in examples/synthetic_demo.py. Run it after a development install with:

python examples/synthetic_demo.py

It prints epoch losses, regression metrics, and a ConstraintReport whose passed=True value verifies the sampled point trajectory. The example is a software demonstration on synthetic data, not a benchmark claim.

Evaluation warning: when several rows belong to one physical condition, the estimator's fallback random validation split is only a convenience for optimization. It is generally not a defensible temporal-extrapolation protocol. Keep condition groups intact, construct the cutoff externally, and pass validation_data= explicitly for scientific evaluation.

Public API

• PriorAnchoredTemporalModel: low-level PyTorch module.
• PriorAnchoredRegressor: fitted preprocessing, training, prediction, and checkpoint wrapper.
• ModelConfig and TrainingConfig: explicit model and optimization settings.
• candidate_time_laws and convex_time_evolution: NumPy implementations of the structured temporal basis.
• audit_trajectory: numerical audit of non-negativity, monotonicity, and optional upper boundedness.
• regression_metrics: R-squared, RMSE, MAE, and MAPE.
• pea_pgnn.concrete: B3-, GL2000-, and ACI209-inspired empirical shrinkage utilities used by the concrete implementation.

PriorAnchoredRegressor.predict_details returns prediction, corrected magnitude and timescale, alpha, normalized weights, the four candidate_laws, their mixed evolution, and the three learned correction terms. These outputs are interpretable model quantities, not automatically identifiable material properties.

Documentation and support

• Method: computational formulation and inherited structural properties.
• Empirical-prior provenance: source formulations, units, compact implementation choices, and applicability limits.
• Data contract: required inputs, shapes, units, and evaluation cautions.
• Model card: intended uses, out-of-scope uses, outputs, evaluation guidance, and risks.
• Research-code map: relationship between this package and the working manuscript code.
• Synthetic example: complete training, prediction, and trajectory-audit workflow.

For bugs, unexpected behavior, or feature requests, open a GitHub issue and include a minimal reproducible example, your Python version, and your PEA-PGNN version. Contributions are welcome; see the contribution guide before submitting a pull request. Security-sensitive reports should follow the private route in the security policy, not a public issue.

Method boundary

The construction guarantees point-prediction properties only when its input contract is respected:

• query time is non-negative and supplied separately from condition context;
• candidate weights are constant with time for a fixed context;
• candidate laws are non-negative, monotone, and bounded on the implemented temporal domain; and
• the learned response magnitude is positive and bounded.

These properties do not automatically extend to prediction-interval endpoints. They also do not establish accuracy, transferability, or validity outside the training domain. See the method documentation and data contract.

The relationship between the cleaned package and the working manuscript code is listed in the research-code map.

Reproducibility scope

This repository intentionally separates reusable method code from research assets. To reproduce the manuscript experiments exactly, a separate archival release should later pin the evaluated dataset version, split identifiers, hyperparameters, trained checkpoints, and figure/table scripts, subject to data licensing and manuscript-publication constraints.

Citation

If this software is useful in your research, please cite the software release using the repository's Cite this repository menu. The metadata are stored in CITATION.cff. The associated manuscript is still being prepared; its final bibliographic citation and DOI will be added when available.

Authors

• Deyu Liang — School of Transportation and Surveying Engineering, Shenyang Jianzhu University, Shenyang, China
• Jinlong Liu — School of Civil Engineering, Southeast University, Nanjing, China
• Lei Xu — Laboratory of Construction Materials, École Polytechnique Fédérale de Lausanne, Lausanne, Switzerland

Acknowledgements

This work was supported by the National Key R&D Program of China (2024YFC38098, 2024YFC3809803); the Liaoning Xingliao Talents Program for Science and Technology Innovation Team of China (No. XLYC2404005); the Technology Research and Development Program of Shenyang Science and Technology Bureau (Grant No. 24-213-3-33).

License

PEA-PGNN is open-source software released under the MIT License. It may be used, copied, modified, distributed, sublicensed, and sold under the terms of that license. The software is provided without warranty.

中文说明

PEA-PGNN 将论文中的知识体系整理为可复用代码：把经验量作为可修正锚点，把多种时间演化规律组成凸组合，并由前向结构保证点预测的非负、单调和有界性质。经验模型来源、单位、简化假设及适用边界见 docs/empirical-priors.md；模型用途、风险和评估要求见 MODEL_CARD.md。

当前仓库是首个干净的软件包版本，包含核心模型、混凝土经验先验、训练封装、约束检查、测试和示例；不包含论文数据库、训练权重、论文正文、审稿材料和实验输出。本项目采用 MIT 开源许可证，作者为 Deyu Liang、Jinlong Liu 和 Lei Xu，资助信息见上方 Acknowledgements。