cwru-sdle / graphfoundationmodels

Code and demos for contructing Data-Driven Digital Twins of Photovoltaic & Advanced Manufacturing systems

graphfoundationmodels

Data-driven Digital Twins: Generalized Foundation Models for Fleet Level System Performance Prediction, Advanced Manufacturing Tool and Part Qualification, Conformance, and Lifetime

Abstract

Digital twins are comprised of computational models that mimic the 'as built' characteristics of devices, systems, and networks of systems whose performance in the real world warrants quantitative and critical assessment. The literature on constructing digital twins is historically focused around task-specific, physics-based models that seek to understand the device from first principles, thereby constructing an idealized understanding of the system. However, these ``empirical digital twins" (eDT) can be quite difficult to construct, as the level of detail required to accurately model the complex physics of many devices is often missing or expensive to obtain, especially for systems with widespread deployment. Additionally, eDTs generally assume the device is working as intended; in practice, many systems experience some form of derating or degradation that causes them to operate off-specification. As it is often the goal of a digital twin model to quantify these departures from the idealized system, it is quite difficult to separate derated assumptions from the expected model output. In contrast, data-driven digital twins (ddDT) seek to model the system as it actually is based on real observations about the device in question. Additionally, ddDTs generally utilize a flexible model architecture (typically an artificial neural network) to avoid injecting implicit bias into the system. This flexibility also lends itself to another advantage: modularity. Digital twins are often utilized to answer multiple different questions about the system. With a ddDT, it is possible to train the model in a self-supervised manner via a reconstruction objective in order to obtain an encoder module. This module can then be used as a Foundation Model (FM) for different tasks without either constructing a entirely new task-specific eDT or training another ddDT from scratch. This work presents a unified pipeline for constructing data-driven Foundation Models for three exemplifying cases: solar photovoltaic fleets, direct ink write additive manufacturing, and laser powder bed fusion. Although these three systems are conceptually very different, the presented Foundation Model utilizes the flexibility of spatiotemporal graph neural networks (st-GNNs) to apply the same methodology to each case, allowing scientists to focus on their scientific objectives rather than troubleshooting a overwhelmingly detailed modeling pipeline.

Downloading data

You can download the data used in this work from OSF with the util module of this package. For example, to download a PV site:

from graphfoundationmodels.util.io import OSF_download
OSF_download(0, 'downloaded_test.csv', './')

graphfoundationmodels's Issues

Jupyter Notebook demo for PV exemplar

A well-written example jupyter notebook with useful Markdown explanations of what is going on needs to be contributed

The PV examplar needs to be less specific to CRADLE

Certain filepaths and other items are hard-coded for our infrastructure. These need to be removed and made general.

The DIW dataloader should be more general

Currently, the DIW dataloader is somewhat hard-coded. The PV dataloader is more flexible, and allows you to specify column names at runtime.

For example, the DIW dataloader currently selects columns to be removed and considered:

        target_cols = ['X_Pos_Error', 'Y_Pos_Error', 'Z_Pos_Error']

        all_cols = df.columns.tolist()
        #remove_cols = ['Time', 'node_id', 'delta_X_Pos', 'delta_Y_Pos', 'delta_Z_Pos']
        remove_cols = ['Time', 'node_id', 'delta_X_Pos', 'delta_Y_Pos',
                       'delta_Z_Pos', 'folder.id', 'layer.num', 'shape', 
                       'on.off', 'velocity', 'z_height', 'acceleration']
        predictor_cols = [col for col in all_cols if col not in target_cols + remove_cols]

A good improvement would be to specify these columns as an input to the dataloader in the init function

On newer versions of pandas, the DIW dataloader fails to pivot_table

On older pandas versions, pivot_table would skip non-numeric columns; this behavior was removed in newer versions. Instead, agg functions must be specified manually.

This concerns the DIW dataloader in graphfoundationmodels.dataloaders.dataloader_DIW.DIWDataset

# throws an error on newer pandas versions
df_pivot = df.pivot_table(index='Timestep', columns='node_id', values=all_cols)

A fix will specify a proper aggregation function like so:

agg_funcs = {col: 'mean' for col in all_cols}  # 'mean' can be changed as required

An automated function to download data from OSF

We are hosting data for this work on OSF at: https://osf.io/dvrc7/

It would be nice if we could write a (python) function to download this data as part of the package so users don't have to do this manually.

See also: cwru-sdle/pvplr#1

Jupyter Notebook for LPBF exemplar

A well-written example jupyter notebook with useful Markdown explanations of what is going on needs to be contributed