News|Articles|October 30, 2025

Testing a New Deep Learning Model for Petroleum Analysis

Listen

0:00 / 0:00

Key Takeaways

IPA, a CNN architecture, improves fuel property prediction accuracy over traditional chemometric and existing deep learning models, especially for small datasets.
The model's ability to learn from raw spectral data without preprocessing is advantageous in the petroleum industry, where data acquisition can be challenging.
IPA demonstrated significant reductions in regression errors compared to PLS and DeepSpectra, enhancing predictive reliability.
Limitations include insufficient testing of model generalizability and challenges in deep learning model interpretability, necessitating further research for broader acceptance.

A recent study presented a novel deep learning model that could improve the prediction of fuel properties in the petroleum industry.

A recent study presented a novel deep learning model that could improve the prediction of fuel properties in the petroleum industry. This study, which was published in the journal Fuel, presents a convolutional neural network (CNN) architecture called “Inception for Petroleum Analysis” (IPA) that outperforms both traditional chemometric methods and existing deep learning benchmarks in the analysis of petroleum samples using near-infrared (NIR) spectroscopy (1).

Why is petroleum analysis important?

Petroleum analysis during upstream, midstream, and downstream production processes is important for several reasons. For one, analysts can examine the fuel for contaminants and remove them before it goes to market (2). Second, it provides an opportunity for the analysis of crude oil fractions (2).

Part of this process is predicting the cetane number (CN) of the middle distillates. The CN is important because it determines how quickly the fuel ignites (3). If the fuel combustion is good, it results in quicker ignition, which is better for car engines (3). For decades, chemometrics, especially partial least squares (PLS) regression, has been the cornerstone of predicting petroleum properties such as the cetane number (CN) of middle distillates. However, the emergence of deep learning has opened new possibilities for more accurate, robust, and faster spectral data interpretation (1).

In this study, the research team tested whether a deep learning framework could not only compete with chemometric models but also surpass them in both predictive accuracy and practical usability for petroleum laboratories (1). The result is “inception for petroleum analysis” (IPA).

What is IPA?

IPA is a CNN architecture inspired by state-of-the-art computer vision systems. Unlike traditional models that rely heavily on preprocessing steps, such as variable selection or digital signal processing, IPA can learn directly from raw spectral data, making it particularly suitable for small datasets, even those containing fewer than 250 samples (1). This benefit is particularly valuable in the petroleum industry, where obtaining large volumes of high-quality data is often difficult or costly.

The study compared IPA’s performance against partial least squares (PLS) and the deep learning benchmark model DeepSpectra across two data sets. The first data set used a reduced spectral range (excluding saturated variables), while the second data set used the full range of spectral features.

What were the results of the study?

The results showed that IPA outperforms existing PLS and deep learning models. On the smaller data set, IPA achieved regression errors 40% lower than PLS, making predictions much closer to measured values (1). When tested on the full-range dataset of 4,150 features, IPA showed a 50% reduction in regression error compared to PLS and was 21% more accurate than DeepSpectra (1).

The researchers determined that what distinguishes IPA from other methods is its unique ability to extract complex representations from spectral data without being misled by regions identified by experts as “saturated,” meaning that IPA filters noise while preserving critical chemical information (1). Ultimately, this task is an aspect to petroleum analysis that traditional chemometric models struggle to achieve without extensive preprocessing (1).

IPA is also simple to use, so there’s less of a technical gap. The architecture uses a limited number of hyperparameters and allows for visual and intuitive fine-tuning, making it accessible for day-to-day petroleum analysis operations (1).

What were the limitations of this study?

In the final part of their article, the researchers acknowledged the limitations of their study. For one, this study did not test model generalizability extensively. Although IPA shows good reproducibility on small data sets, there are many petroleum sources and conditions, most of which were unexplored in this study (1). Additionally, the interpretability of deep learning models continues to be a challenge. As the authors note, few studies have yet delved deeply into explaining how deep learning models make their spectral predictions, which is a critical step toward broader industry acceptance (1).

However, the researchers did lay the groundwork for future work in NIR spectroscopy in hydrocarbon analysis. The IPA model helps bridge the gap between deep learning and traditional chemometric approaches. It provides a robust framework that minimizes the need for manual preprocessing while enhancing predictive reliability (1). Because the energy industry requires precision and efficiency as these variables directly influence fuel quality and regulatory compliance, innovations such as the IPA model could redefine analytical standards (1).

As a result, the IPA method can help advance the accuracy, efficiency, and interpretability of fuel characterization, which marks a new innovation in data-driven petroleum science.