Literature Derived Polypropylene Mechanical Property Data for an Automotive Material Development Case Study Using a Quality Controlled Open Access Extraction Framework

Science has long promised a future where new materials are designed by computers, sifting through vast oceans of data to find the perfect combination of strength, weight, and flexibility. For decades, researchers have believed that the answers were already written down, scattered across millions of scientific papers. The challenge was never a lack of information, but rather the difficulty of turning those written descriptions into a format a computer could actually use. A scientist reading a paper can easily connect a specific plastic recipe to a test result, but a computer sees only words and numbers without context. To build a machine that can learn from these documents, the data must be extracted with extreme care, linking every single number back to the exact sentence and table where it was found, ensuring that the story behind the number is never lost.

This is the precise problem tackled by a team of researchers at Széchenyi István University, who set out to build a system capable of turning open-access scientific articles into a reliable, traceable database for materials science. They focused their efforts on polypropylene, a common plastic used in everything from packaging to car parts, specifically looking for data on how adding different ingredients changes its mechanical strength. The researchers did not simply ask a computer to read the papers and guess the answers. Instead, they constructed a multi-layered workflow that acts as a rigorous filter. First, the system finds the documents and converts them from their original PDF format into a structured digital text. Then, it breaks these texts into small, manageable windows of evidence. Finally, it uses artificial intelligence to identify potential data points, but with a crucial twist: the system is designed to admit when it is unsure. It separates the raw evidence found in the text from the computer's best guess at what that evidence means, creating a clear path for human experts to review the most promising findings.

The team tested this framework by processing one hundred valid scientific papers on polypropylene. The system successfully converted these documents into thousands of small evidence windows, capturing the specific sections where data was discussed. From this massive pool, the computer generated nearly nine hundred candidate records, each representing a potential link between a material recipe and a measured property. However, the true value of the study lies not in how much data it found, but in how strictly it judged that data. When the first round of automated checks was applied, the system kept only eighty-four candidates as high-quality records, flagged sixty-six for human review, and rejected the remaining seven hundred and forty-nine. A second, more detailed check confirmed that the vast majority of the initial candidates were missing critical details, such as the specific amount of an additive or the exact conditions under which a test was run. In the end, only ninety-one of the original candidates contained all the necessary information in a syntactically complete format, while a stricter policy requiring simultaneous support for modifiers, loading, property, value, unit, and the formulation-property relation retained just seventy-six candidates.

The researchers found that while finding the right documents and preserving their origin is relatively straightforward, reconstructing the full story of a material experiment is incredibly difficult. The system often struggled to connect a specific number to the correct recipe because the information was scattered across different parts of a table or hidden in footnotes. For instance, a table might list a strength value without immediately stating which mixture of plastic and fiber produced it, requiring the reader to look back at the column headers or surrounding text. The study showed that even advanced artificial intelligence models, when asked to extract this information, frequently had to admit they could not find a clear link. In one specific check, the system identified that the relationship between a formulation and a property was not explicit in over four hundred candidates, and in more than three hundred cases, the amount of a modifier was simply missing. These gaps meant that the computer could not confidently treat those records as ready for analysis.

Despite these challenges, the framework proved its worth by creating a transparent infrastructure where every single piece of data remains connected to its source. The researchers built two different ways to access this information. One method allows users to search through the raw evidence windows, seeing exactly where a number came from in the original text. The other method creates a compressed "wiki" of facts, a smaller, faster index that summarizes what the project has learned so far about recurring material classes and unresolved gaps. This dual approach ensures that while the system can quickly point researchers toward relevant papers, it never hides the underlying evidence. The study explicitly states that the results are not a final, perfect dataset ready for immediate industrial use, but rather a sophisticated screening tool. It successfully identified which families of plastic formulations are worth further investigation, but it stops short of making final recommendations for car parts or other applications.

The ultimate goal of this work is to shift the burden of data collection from a manual, error-prone task to a structured, auditable process. By separating the raw evidence from the machine-generated candidates, the system allows human experts to focus their time on the most uncertain and valuable records. The researchers demonstrated that with the right quality controls, it is possible to turn a chaotic collection of scientific papers into a navigable map of knowledge. They found that while the computer can do the heavy lifting of finding and organizing the text, the final step of confirming that a specific property belongs to a specific material recipe still requires human judgment. The study concludes that the path forward is not simply to download more papers, but to build expert-verified reference sets that can teach the system to recognize these complex relationships with greater accuracy. Until that happens, the system serves as a powerful guide, pointing scientists toward the most promising leads while keeping a clear record of what is known, what is suspected, and what remains to be discovered.