Method and system for extracting and converting coating composition information in a material document

Abstract

The application discloses a kind of material literature coating composition information extraction conversion method and system, including input processing module, image recognition module, rule extraction module, intelligent extraction module, composition conversion module, data check module, batch management module and file export module, can realize the multimodal intelligent extraction of alloy composition, high-precision bidirectional conversion of atomic percentage (at.%) and mass percentage (wt.%), batch data batch management.The application can convert coating composition in unstructured or semi-structured material literature into structured data that can be calculated, checked and exported, improving material composition data arrangement efficiency and consistency.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology and relates to a method and system for extracting and converting coating composition information from materials literature. Background Technology

[0002] In research on high-entropy alloy coatings, laser cladding coatings, and wear- and corrosion-resistant coatings, the literature typically describes the coating composition in various ways. For example, some literature directly provides explicit percentages such as Co 20 wt.% and Cr 20 wt.%; some literature uses natural language or chemical formulas to describe it, such as AlCoCrFeNi equimolar alloy or CoCrFeNiMn molar high-entropy alloy coating; and some literature uses variable subscripts or numerical subscripts such as CoCrFeNiMox, CoCrMnNiMo0.2, and CoCrMnNiMo_0.1. For coatings containing additive phases, expressions such as CoCrFeNi-10WC, (CoCrFeNi)90(WC)10, and base alloy with 10 wt.% WC are also commonly used.

[0003] Current data processing methods typically require researchers to read papers, screenshots, or tables, manually determine elements, units, proportion types, and sample numbers, and then enter the results into a spreadsheet. This process is inefficient, has poor repeatability, and is prone to omissions or misinterpretations. Especially when converting between mass percentage and atomic percentage, calculations based on element atomic weights are necessary, and manual calculations are susceptible to rounding errors or unit confusion.

[0004] While existing general text extraction or large language model question answering methods can recognize some natural language content, they lack stable rule constraints on equimolarity, variable subscripts, balance elements, added phase coatings, and unit conversions in material chemical formulas, which can easily lead to problems such as inconsistent formats, opaque calculations, or unverifiable results. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method and system for extracting and converting coating composition information from materials literature. This method combines the determinism of rule-based extraction with the generalization ability of intelligent extraction, thereby generating structured data that can be processed in batches, verified, and exported.

[0006] This invention is achieved through the following technical solution: A method for extracting and converting coating composition information from materials literature includes the following operations: S1. The input processing module acquires the literature data of the material to be processed, which includes text data, image data and sample name; S2. When the literature data to be processed includes image data, the image recognition module performs format verification, image preprocessing, and optical character recognition on the image data to obtain the recognized text. S3. The text data and the identified text are normalized to obtain the text to be extracted; and the coating representation and candidate element symbols are extracted from the text to be extracted and the sample name. S4. The rule extraction module sequentially performs phase coating extraction, explicit component extraction, equimolar component extraction, and variable subscript component extraction on the text to be extracted. The extraction of the added phase coating involves detecting whether it is an added phase coating and identifying the matrix alloy, the chemical formula of the added phase, the content of the added phase, and the unit. Clearly define the components: identify the correspondence between element symbols and numerical content, and make up for the remaining amount to ensure that the content of each element is logically closed. Equimolar extraction: After detecting keywords including equimolar / equiatomic ratio, the effective element set is extracted from the continuous element markers, and the content of each element is normalized to a percentage under the same benchmark. Variable subscript extraction: Parse the subscript variables and their values, and normalize the content of each element into a percentage under a unified standard according to a predetermined proportional relationship; If the rule extraction module does not obtain effective components, or is forced to extract intelligent components, the intelligent extraction module calls the large language model to extract intelligent components from the text to be extracted. S5. Generate coating component data based on the component types extracted by the rule extraction module or intelligent extraction module; the coating component data includes coating expression, element symbols, element content, content unit, extraction mode, extraction status and status description; S6. The component conversion module performs bidirectional conversion between mass percentage and atomic percentage of the coating component data based on the element atomic weight data. S7. The data verification module performs element validity verification, total component verification, and single element abnormality percentage verification on the coating component data. S8, the batch management module saves the verified coating composition data by batch; the file export module generates exportable structured files and exports them.

[0007] Furthermore, the image preprocessing and optical character recognition include: Verify image format and file size; Convert the image to a uniform color mode; When the image size is below a preset threshold, the image is magnified. The recognized text is obtained using an optical character recognition engine that supports mixed Chinese and English recognition; The identified text is then processed to perform whitespace character normalization and common recognition error correction. The normalization of the text data and the identified text includes: First, perform basic normalization, including removing leading and trailing whitespace, to form the string to be parsed; Then, based on the alloy / coating marking pattern in the sample name or text, extract or complete the coating marking information.

[0008] Furthermore, the specific extraction of the defined components includes the following operations: Use pattern matching to extract element symbols and content, and standardize the element symbols into a standard chemical nomenclature format; Identify at least one unit expression among mass percentage, atomic percentage, mass fraction, atomic fraction, and their English expressions, and standardize the units to a unified standard; Identify surplus element keywords, wherein the surplus element keywords include at least one of Bal., balance, surplus, remainder, or surplus; When there is a surplus element and the total known element content is less than 100, the surplus value is obtained by subtracting the total known element content from 100, and the surplus value is allocated to the surplus element.

[0009] Furthermore, the extraction of equimolar components specifically includes the following operations: The text to be extracted was tested for keywords including equimolar, equiatomic, equal molar, equal atomic ratio, and equiatomic. Multiple element symbols are parsed from the coating representation or the text to be extracted; The atomic percentage of each element is calculated by dividing 100 by the number of elements, and the content unit is set to atomic percentage.

[0010] Furthermore, the extraction of variable subscript components specifically includes the following operations: Identify coating representations containing variable subscripts, coating representations containing numerical subscripts, or coating representations containing underlined numerical values; Determine matrix elements, variable elements, and variable values ​​from the coating representation, the text to be extracted, and / or the sample name; When the variable value is x and the number of matrix elements is n, the atomic percentage of each matrix element is calculated as 100 / (n+x), and the atomic percentage of the variable element is calculated as x×100 / (n+x); When the same variable subscript component contains multiple variable values, an independent coating component data is generated for each variable value, and the matrix element, variable element, variable value and subscript mode type are recorded in the coating component data.

[0011] Furthermore, intelligent component extraction using a large language model includes the following operations: Construct prompt information based on the text to be extracted, the sample name, and / or the image context; The instruction is to output the coating representation, element list, content unit, extraction mode, explicit component information, subscript information, and confidence level in a pre-defined JSON structure using the large language model. Perform JSON parsing, field validation, element symbol validation, and component calculation on the content returned by the large language model; When the large language model returns equimolar components or variable subscript components, the element content is calculated.

[0012] Furthermore, the bidirectional conversion between the mass percentage and the atomic percentage includes the following operations: When converting from mass percentage to atomic percentage, it is calculated according to the formula at_i=(wt_i / M_i) / Σ(wt_j / M_j)×100, where wt_i is the mass percentage of the i-th element and M_i is the atomic weight of the i-th element; When converting from atomic percentage to mass percentage, it is calculated according to the formula wt_i=(at_i×M_i) / Σ(at_j×M_j)×100, where at_i is the atomic percentage of the i-th element; After conversion, the element content will be formatted to a uniform precision, and the content unit will be updated.

[0013] Furthermore, the element validity verification, total component verification, and single element abnormality percentage verification include the following operations: Match element symbols with a preset element table; Determine whether the sum of the extracted component data falls within the first threshold range; Determine whether the sum of the components obtained from the conversion is within the second threshold range; A warning is generated when the content of a single element exceeds a preset threshold for that single element. The error messages, warning messages, and verification details are output in association with the coating composition data.

[0014] Furthermore, the step of saving in batches and generating exportable structured files includes the following operations: Assign incremental batch numbers to one or more extraction or conversion operations; Record the creation time and coating quantity for each batch; Summarize the element symbols appearing in all batches as a dynamic header; Write the batch number, coating expression, extraction mode, extraction status, status description, content of each element and content unit into an electronic spreadsheet file.

[0015] This invention also provides a system for extracting and converting coating composition information from materials literature, comprising: The input processing module acquires text data, image data, and / or sample names; The image recognition module performs format verification, preprocessing, and optical character recognition on image data. The rule extraction module performs phase coating detection, explicit component extraction, equimolar component extraction, and variable subscript component extraction on the recognized text. The extraction of the added phase coating involves detecting whether it is an added phase coating and identifying the matrix alloy, the chemical formula of the added phase, the content of the added phase, and the unit. Clearly define the components: identify the correspondence between element symbols and numerical content, and make up for the remaining amount to ensure that the content of each element is logically closed. Equimolar extraction: After detecting keywords including equimolar / equiatomic ratio, the effective element set is extracted from the continuous element markers, and the content of each element is normalized to a percentage under the same benchmark. Variable subscript extraction: Parse the subscript variables and their values, and normalize the content of each element into a percentage under a unified standard according to a predetermined proportional relationship; The intelligent extraction module calls a large language model to perform intelligent component extraction on the text to be extracted when the rule extraction fails to obtain effective components or when forced intelligent extraction is specified. The composition conversion module performs bidirectional conversion between mass percentage and atomic percentage based on element atomic weight data. The data verification module performs element validity verification, component sum verification, and single element abnormality percentage verification. The batch management module saves coating composition data for different batches; The file export module generates a structured file containing batch information, extraction status, and element content.

[0016] Compared with the prior art, the present invention has the following beneficial technical effects: This invention can classify and process various coating component expressions commonly found in materials literature, thereby improving extraction accuracy. Through a process of "rule priority and intelligent supplementation," it ensures deterministic calculation while taking into account the generalization and recognition capabilities of natural language paragraphs. It performs explicit calculations on residual elements, equimolar components, and variable subscript components, avoiding the generation of untraceable results solely based on natural language models.

[0017] This invention supports OCR image input and can process non-pure text materials such as paper screenshots and table screenshots; this invention supports bidirectional conversion between mass percentage and atomic percentage, and improves data reliability through element validity and sum verification; this invention supports batch management and spreadsheet export, which facilitates subsequent material database construction, statistical analysis, and experimental result verification.

[0018] The present invention also includes a component conversion module and a component verification module. The component conversion module, based on the standard atomic weight table of elements, calculates the molar contribution of each element by dividing the mass fraction by the atomic weight and normalizes it to an atomic percentage, or calculates the mass contribution of each element by multiplying the atomic fraction by the atomic weight and normalizes it to a mass percentage, thereby realizing bidirectional conversion between mass percentage and atomic percentage. For the balance element, the difference between the sum of the known element contents and 100% is assigned to the balance element before conversion.

[0019] The component verification module performs element validity checks based on a valid set of element symbols consistent with the atomic weight table. For the extraction results, warnings are generated when the total component is outside the first preset range (e.g., 85%–105%), the number of elements is too small, or the proportion of a single element exceeds the first threshold. Errors are generated when invalid element symbols are present. For the conversion results, warnings are generated when the total component is outside the second preset range (e.g., 99%–101%), and element validity checks are performed again. After completing the extraction verification, the system performs unit conversion and performs post-conversion verification on the conversion results. The verified coating data is then saved to the batch and can be exported, thereby improving the reliability and traceability of material literature component data. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the coating composition information extraction and conversion system in materials literature according to the present invention; Figure 2 This is a flowchart of the method for extracting and converting coating composition information from materials literature according to the present invention; Figure 3 This is a flowchart illustrating the conversion and verification process between mass percentage and atomic percentage in this invention. Figure 4 This is a flowchart illustrating the multi-source input processing of text, images, and intelligent recognition in this invention. Figure 5 This is one example of the coating output of the present invention; Figure 6 This is a second example of the coating output of the present invention; Figure 7 This is the third example of the coating output of the present invention; Figure 8 This is the fourth example of the coating output of the present invention. Detailed Implementation

[0021] This invention discloses a method and system for extracting and converting coating composition information from materials literature. The method adopts a hybrid architecture of rule engine + large language model (LLM), which can realize multimodal intelligent extraction of alloy composition, high-precision bidirectional conversion of atomic percentage (at.%) and mass percentage (wt.%), and batch management of batch data. It completely solves the core pain points of the industry such as low efficiency, easy error and poor format adaptability of traditional manual processing of literature composition data.

[0022] See Figures 1-4 A method for extracting and converting coating composition information from materials literature, comprising the following operations: S1. The input processing module acquires the literature data of the material to be processed, which includes text data, image data and sample name; S2. When the literature data to be processed includes image data, the image recognition module performs format verification, image preprocessing, and optical character recognition on the image data to obtain the recognized text. S3. The text data and the identified text are normalized to obtain the text to be extracted; and the coating representation and candidate element symbols are extracted from the text to be extracted and the sample name. S4. The rule extraction module sequentially performs phase coating extraction, explicit component extraction, equimolar component extraction, and variable subscript component extraction on the text to be extracted. The extraction of the added phase coating involves detecting whether it is an added phase coating and identifying the matrix alloy, the chemical formula of the added phase, the content of the added phase, and the unit. Clearly define the components: identify the correspondence between element symbols and numerical content, and make up for the remaining amount to ensure that the content of each element is logically closed. Equimolar extraction: After detecting keywords including equimolar / equiatomic ratio, the effective element set is extracted from the continuous element markers, and the content of each element is normalized to a percentage under the same benchmark. Variable subscript extraction: Parse the subscript variables and their values, and normalize the content of each element into a percentage under a unified standard according to a predetermined proportional relationship; If the rule extraction module does not obtain effective components, or is forced to extract intelligent components, the intelligent extraction module calls the large language model to extract intelligent components from the text to be extracted. S5. Generate coating component data based on the component types extracted by the rule extraction module or intelligent extraction module; the coating component data includes coating expression, element symbols, element content, content unit, extraction mode, extraction status and status description; S6. The component conversion module performs bidirectional conversion between mass percentage and atomic percentage of the coating component data based on the element atomic weight data. S7. The data verification module performs element validity verification, total component verification, and single element abnormality percentage verification on the coating component data. S8, the batch management module saves the verified coating composition data by batch; the file export module generates exportable structured files and exports them.

[0023] This invention also provides a system for extracting and converting coating composition information from materials literature, comprising: The input processing module acquires text data, image data, and / or sample names; The image recognition module performs format verification, preprocessing, and optical character recognition on image data. The rule extraction module performs phase coating detection, explicit component extraction, equimolar component extraction, and variable subscript component extraction on the recognized text. The intelligent extraction module calls a large language model to extract coating component information when the rule extraction fails to obtain effective components or when forced intelligent extraction is specified. The composition conversion module performs bidirectional conversion between mass percentage and atomic percentage based on element atomic weight data. The data verification module performs element validity verification, component sum verification, and single element abnormality percentage verification. The batch management module saves coating composition data for different batches; The file export module generates a structured file containing batch information, extraction status, and element content.

[0024] The following provides a detailed explanation of each operation step and module.

[0025] The input processing module receives text, image files, sample names, and target unit parameters submitted by the user; the text can be paragraphs from a paper, copied text from a table, experimental record text, or a mixture of Chinese and English content; the image can be a screenshot containing a composition table or text description.

[0026] The image recognition module performs image verification and OCR recognition. Verification includes checking if the file is empty, if the file size exceeds a preset limit, and if the image format is a supported format such as PNG, JPG, JPEG, GIF, BMP, or WebP. Image preprocessing includes converting the image to a unified color mode and enlarging it if the image size is small. Finally, an OCR engine that supports mixed Chinese and English text recognition generates the recognized text.

[0027] Specifically, the image preprocessing and optical character recognition include: Verify image format and file size; Convert the image to a uniform color mode; When the image size is below a preset threshold, the image is magnified. The recognized text is obtained using an optical character recognition engine that supports mixed Chinese and English recognition; The identified text is then processed to perform whitespace character normalization and common recognition error correction. The normalization of the text data and the identified text includes: First, perform basic normalization, including removing leading and trailing whitespace, to form the string to be parsed; Then, based on the alloy / coating marking pattern in the sample name or text, extract or complete the coating marking information.

[0028] The extraction of candidate coating representations and candidate element symbols involves identifying chemical formulas composed of multiple consecutive element symbols, as well as individually listed element symbols, units, and percentage values.

[0029] The rule extraction module (rule engine) is responsible for deterministic extraction. Internally, it may include a phase detection unit, a specific component extraction unit, an equimolar component calculation unit, and a variable subscript resolution unit. The rule extraction module extracts rules in order of priority: First, the additive phase coating is extracted, and it is detected whether it is an additive phase coating. The matrix alloy, chemical formula of the additive phase, content and unit of the additive phase are identified. For additive phase coatings that do not yet support direct separation and extraction, the additive phase state data is generated and the original information is retained.

[0030] Second, clarify the extraction of components, identify the elements and their content values, mass percentages or atomic percentages, and handle residual elements such as Bal., balance, and balance.

[0031] Third, equimolar component extraction: when keywords such as equimolar, equiatomic ratio, and equimolar are detected, the atomic percentage of each element is calculated based on the number of elements.

[0032] Fourth, extract the components by variable subscripts, analyze the expressions such as CoCrFeNiMox, CoCrMnNiMo0.2, and CoCrMnNiMo_0.1, and calculate the atomic percentage of the matrix element and the variable element based on the variable values.

[0033] The rule extraction module operates according to a preset priority to reduce misjudgments between different modes. The addition phase detection, explicit component extraction, equimolar identification, and variable subscript parsing within the rule extraction module logically represent the multi-class recognition capabilities of the module. In practice, these are preferably executed sequentially according to the aforementioned priority order to avoid cross-misjudgments between different modes.

[0034] The intelligent extraction module is responsible for handling situations where natural language expressions are complex, rule extraction fails, or users request mandatory intelligent processing.

[0035] The rule extraction module and the intelligent extraction module adopt a hierarchical processing strategy of "rule priority, intelligent supplementation". The rule extraction module and the intelligent extraction module are connected through a scheduling unit. The scheduling unit performs sequential rule matching on the preprocessed unified text to be extracted and calls the intelligent extraction module when the triggering conditions are met.

[0036] After generating the rule extraction results, the system determines whether the rule results are valid: if at least one of the following is successfully identified and the corresponding rule extraction results are generated, the rule results are deemed valid and the intelligent extraction module is not invoked; if none of the above four items are successfully identified, the rule results are deemed invalid and the intelligent component extraction step is entered.

[0037] When rule extraction fails to yield valid results, or when the user specifies mandatory intelligent extraction, the large language model is invoked for intelligent component extraction. During intelligent extraction, a prompt message is constructed containing the text to be extracted, the sample name, and the image context. The large language model is then requested to return the coating representation, element list, units, extraction mode, explicit component information, subscript information, and confidence level according to a preset JSON structure. The returned results are then parsed using JSON and validated, and in scenarios involving equimolarity or variable subscripts, the element content is recalculated using local rules, thereby reducing the uncertainty introduced by free text output.

[0038] The intelligent extraction module constructs prompt words through an interface compatible with large language models, requires the model to return structured JSON, and performs secondary verification and calculation locally by the system.

[0039] The large language model can be based on OpenAI, Python, SDK, or other models. As long as the third-party large model service is compatible with the OpenAI Chat Completions API, it can be accessed by configuring api_key, base_url, and model.

[0040] Furthermore, the construction method of the prompt words is divided into two parts; the first part is the fixed system prompt words, which tell the model: task, output, etc. For example, the following prompts: You are a materials science assistant.

[0041] The task is to extract the components of high-entropy alloys / coatings.

[0042] The element, content, and unit type must be identified.

[0043] wt.% is mapped to the mass percentage as basis: "wt.%".

[0044] at.%, atomic percentage, and molar ratio are mapped to basis: "at.%".

[0045] Do not perform unit conversion; the backend will handle it.

[0046] The output must be in a fixed JSON format, containing fields such as coatings, notation, elements, extraction_mode, explicit_composition, basis, and confidence.

[0047] The second part is dynamic user suggestions, which are dynamically concatenated based on the input: Python if has_image: prompt_parts.append("Please extract the coating / alloy composition information from the image and text:\n") else: prompt_parts.append("Please extract the coating / alloy composition information from the following text:\n") if sample_name: prompt_parts.append(f"Sample name: {sample_name}\n") if text and text.strip(): prompt_parts.append(f"Text content:\n{text}\n") If a variable index like x=0.1, 0.2 is detected, an additional reminder will be added: "Note: Multiple variable values ​​have been detected. Please calculate and generate an independent coating entry for each value separately." Finally, the constraint "Please output according to the specified JSON format" will be added.

[0048] The coating composition data includes the coating expression, element list, content unit, extraction mode, extraction status, status description, and optional subscript information. The system uses a preset element table to verify the validity of element symbols and performs component sum verification on the extraction and conversion results respectively; when the number of elements is too small, the proportion of a single element is too high, or the sum exceeds the threshold, a warning is output; when the element symbol is invalid or the result is empty, an error is output.

[0049] When a user requests unit conversion, a bidirectional conversion between wt.% and at.% is performed based on the elemental atomic weight data. When converting mass percentage to atomic percentage, the molar contribution is first obtained by dividing the mass percentage by the atomic weight, and then normalized to a percentage. Conversely, when converting atomic percentage to mass percentage, the mass contribution is first obtained by multiplying the atomic percentage by the atomic weight, and then normalized to a percentage. After conversion, the coating data is saved to a batch and exported as a spreadsheet file.

[0050] The composition conversion module uses an element atomic weight table to perform mass percentage to atomic percentage conversion. Preferably, the element atomic weights use a standard atomic weight data table, and when an unknown element is encountered during the conversion process, the conversion for that element is not performed and an error or warning is given.

[0051] The data verification module is used to determine the validity of element symbols, the number of elements, the total composition, and the percentage of abnormal single elements.

[0052] Specifically, the verification of element validity, the verification of total composition, and the verification of abnormal proportion of single elements include the following operations: Match element symbols with a preset element table; Determine whether the sum of the extracted component data falls within the first threshold range; Determine whether the sum of the components obtained from the conversion is within the second threshold range; A warning is generated when the content of a single element exceeds a preset threshold for that single element. The error messages, warning messages, and verification details are output in association with the coating composition data.

[0053] The batch management module assigns a batch number to each save operation and records the time. The file export module summarizes all batch data and generates a spreadsheet.

[0054] Specific implementation examples are given below.

[0055] Example 1: Acquisition of Recognized Text: The input processing module acquires at least the following literature data for the material to be processed: text data entered or pasted by the user, recognized text obtained through optical character recognition, and sample name as auxiliary context; among which, the sample name is mainly used for context parsing of coating expression, variable subscript and intelligent extraction prompts, and is generally not incorporated into the main text string of the text to be extracted.

[0056] It should be noted that the target unit parameters submitted by the user (such as mass percentage wt.% or atomic percentage at.%) are inputs for the composition conversion stage and are not within the scope of the text to be extracted.

[0057] After receiving the text, image file, sample name, and target unit parameters submitted by the user, the input processing module generates the text to be extracted in the following manner: (1) Text data path When the material literature data includes text data, the beginning and end blanks of the paper paragraphs, table copied text, experimental records or mixed Chinese and English content submitted by the user are straightened, and the straightened text is used as the text to be extracted.

[0058] (2) Image data path When the material document data includes image data, the image recognition module performs preprocessing on the image, such as format verification, color mode conversion, and size enlargement, and obtains the recognized text through an optical character recognition engine that supports mixed Chinese and English recognition; then, the recognized text is normalized for whitespace characters and corrected for common recognition errors, and the corrected recognized text is used as the text to be extracted.

[0059] (3) Multi-source combined path When both user text descriptions and image data exist, the user text descriptions and recognized text can be merged according to a preset format to form a unified text to be extracted that includes both document descriptions and image recognition content, for use by the intelligent extraction module.

[0060] (4) Image visual recognition path (optional) When using a large language model with visual understanding capabilities, image content can directly participate in intelligent extraction as image context; in this case, the text to be extracted still refers to the text portion of the prompt information, while image information is provided separately in the form of multimodal input.

[0061] Preprocessing refers to performing file size and format checks on image data; after decoding the verified image into a bitmap, if its color mode is not RGB, it is converted to the RGB unified color mode; if the image width or height is lower than the preset minimum side length threshold (such as 800 pixels), the image is enlarged to above the threshold by using Lanczos interpolation while maintaining the aspect ratio.

[0062] After completing the pixel preprocessing described above, an optical character recognition engine that supports mixed recognition of simplified Chinese and English is invoked, and a unified text block page segmentation mode is used for recognition. After recognition, blank lines are removed, inline whitespace is normalized, and common misrecognition corrections for percentage signs and element symbols are performed on the recognized text to obtain the recognized text for subsequent component extraction.

[0063] Specifically: If the original image is not RGB (such as PNG's RGBA with an alpha channel, P palette mode, L grayscale, CMYK, etc.), it will be uniformly converted to RGB three channels. The purpose is to avoid inconsistencies in how Tesseract / Pilow handles non-scale images; transparent areas, indexed colors, etc., are standardized to a format that is easier for the OCR engine to process before entering the OCR process.

[0064] Identification text refers to a unified text carrier obtained after the acquisition, identification, and standardization of the literature data of the material to be processed, which is used for subsequent extraction of coating composition information.

[0065] Example 2: Rule Extraction and Level Recognition Added phase detection identifies composite coatings described as "matrix alloy + added phase compound + added phase content" and determines them to be added phase coatings. Instead of expanding them into a table of atomic / mass fractions of each element, it records structured auxiliary information and marks "full component extraction is not currently supported".

[0066] Typical recognition format: CoCrFeNi-10WC, AlCoCrFeNi+15TiC; (CoCrFeNi)90(WC)10; CoCrFeNi with 10 wt.% WC; Output requirements: Please provide the output results in the above recognition format; the results should look like this: Figure 5 As shown.

[0067] The system clearly extracts components, directly extracting the given element symbols and numerical content from the text, and identifying wt.% / at.% (and Chinese and English unit aliases); it supports Bal. / remaining amount, etc., and calculates the remaining amount as 100% - the sum of known elements.

[0068] Typical input: Co 20%, Cr 20%, Fe 20%, Ni 20%, Mn 20%, wt.% Al 10 wt.%, Co 22.5 wt.%, Cr 22.5 wt.%, Fe Bal. Output requirements: Please provide the output results for the typical input above; the results are as follows: Figure 6 As shown.

[0069] Equimolar extraction: When the literature specifies the equimolar / equiatomic ratio (Chinese and English keywords, such as "equimolar", "equimolar", "equal atomic", etc.), but does not give the specific percentage of each element, the element set is parsed from the chemical formula string, and the atomic fraction of each element is calculated deterministically according to at.% = 100 / n (n is the number of elements).

[0070] Typical input: Equimolar high-entropy alloy coating of CoCrFeNiMn; AlCoCrFeNi equimolar alloy; Key points for output are as follows Figure 6 As shown.

[0071] Variable subscript extraction handles expressions where the content of a certain element is represented by the subscript variable x or a specific decimal; it parses the matrix elements, variable elements, and the value of x (through pattern matching such as x=0.1, 0.2, direct subscript Mo0.2, underscore Mo_0.1, etc.), and then calculates at.% using a fixed formula: Matrix element: 100 / (n + x) Variable element: x × 100 / (n + x) (n is the number of base elements) Typical input: CoCrFeNiMox (x=0.2, molar ratio) CoCrMnNiMo0.2.

[0072] Preset priority and examples The rule module strictly follows the order and returns a result as soon as a match is found: Priority 1 → Add phase detection unit Priority 2 → Define the component extraction unit Priority 3 → Equimolar composition calculation unit Priority 4 → Variable subscript resolution unit (After rule failure) → Smart extraction (non-rule module) Priority examples are shown in Table 1.

[0073] Table 1. Examples of priority for rule modules The purpose of this design is to prioritize the addition phase. If the specific composition of CoCrFeNi-10WC is analyzed first, only some numbers or incorrect units may be resolved. Identifying the composite structure first can avoid treating the addition phase coating as a regular HEA composition table.

[0074] The composition must be clearly stated before the equimolar composition: Co 20%, Cr 20%... The specific values ​​have been given and should not be overshadowed by the occasional appearance of words such as "equal" in the text.

[0075] Equimolarity precedes subscript: Equimolarity of Co, Cr, Fe, Ni, and Mn does not have an x ​​subscript and should be processed by equimolarity units first; avoid mistakenly subscripting consecutive elements into subscript chemical formulas.

[0076] Subscripts are placed at the end: CoCrMnNiMo0.2 is similar to pure element strings. It is used for special analysis when it is clear that equimolarity is not applicable, so as to reduce conflicts with ordinary chemical formulas. This priority strategy can reduce the probability of the same text being parsed repeatedly or incorrectly by units corresponding to multiple expression types.

[0077] Without prioritization, the same text might be "misidentified" by multiple units; for example, CoCrFeNi-10WC looks like both an alloy name and a string of numbers. Trying different expressions in sequence according to the preset priority is to reduce misjudgment between different expression modes.

[0078] Example 3: The extraction of specific components includes the following steps: Use pattern matching to extract element symbols and content, and standardize the element symbols into a standard chemical nomenclature format; Identify at least one unit expression among mass percentage, atomic percentage, mass fraction, atomic fraction, and their English expressions, and standardize the units to a unified standard; Identify surplus element keywords, wherein the surplus element keywords include at least one of Bal., balance, surplus, remainder, or surplus; When there is a surplus element and the total known element content is less than 100, the surplus value is obtained by subtracting the total known element content from 100, and the surplus value is allocated to the surplus element.

[0079] For example, when the input text is "Co 20%, Cr 20%, Fe 20%, Ni 20%, Mn 20%, wt.%", the system first detects that the unit is expressed as a mass percentage. It then identifies the elements Co, Cr, Fe, Ni, and Mn, and their corresponding contents. The system generates coating composition data, with the extraction mode set to explicit composition mode and the extraction status showing extracted components with units in wt.%.

[0080] When the input text is “Al 10 wt.%, Co 22.5 wt.%, Cr 22.5 wt.%, Fe Bal.”, the system identifies Fe as the balance element. Given that the total element content is 55, the Fe content is calculated to be 45. The system outputs Al 10, Co 22.5, Cr 22.5, Fe 45, in wt.%.

[0081] After the components are extracted, the system checks the validity of the element symbols and determines whether the sum of the components is within the extraction threshold range. If the sum deviates significantly from the preset range, a warning message is added to the result; however, if there are no element errors, the result can still be retained for user review.

[0082] Example 4: Equimolar component extraction includes the following operations: The text to be extracted was tested for keywords including equimolar, equiatomic, equal molar, equal atomic ratio, and equiatomic. Multiple element symbols are parsed from the coating representation or the text to be extracted; The atomic percentage of each element is calculated by dividing 100 by the number of elements, and the content unit is set to atomic percentage.

[0083] For example, when the input text is "CoCrFeNiMn equimolar high-entropy alloy coating", the system detects the keyword "equimolar" and parses the five elements Co, Cr, Fe, Ni, and Mn from the coating expression. Since the number of elements n is 5, the system calculates the atomic percentage of each element as 100 / 5 = 20, and the output unit is at.%.

[0084] When the input text is “AlCoCrFeNi equimolar alloy”, the system detects the English keyword “equimolar” and automatically calculates the atomic percentage based on the number of elements.

[0085] The above processing method enables the conversion of material literature expressions that do not explicitly list percentage values ​​but have equimolar semantics into structured component data. Example 5: Variable subscript component extraction includes the following operations: Identify coating representations containing variable subscripts, coating representations containing numerical subscripts, or coating representations containing underlined numerical values; Determine matrix elements, variable elements, and variable values ​​from the coating representation, the text to be extracted, and / or the sample name; When the variable value is x and the number of matrix elements is n, the atomic percentage of each matrix element is calculated as 100 / (n+x), and the atomic percentage of the variable element is calculated as x×100 / (n+x); When the same variable subscript component contains multiple variable values, an independent coating component data is generated for each variable value, and the matrix element, variable element, variable value and subscript mode type are recorded in the coating component data.

[0086] For example, when the input text is “CoCrFeNiMox (x=0.2, molar ratio)”, the system identifies Co, Cr, Fe, and Ni as matrix elements, and Mo as a variable element with a value of 0.2 for x. The number of matrix elements, n, is 4, therefore the total molar ratio is 4.2. The system calculates the atomic percentage of each matrix element as 100 / 4.2, and the atomic percentage of Mo as 0.2×100 / 4.2, outputting the unit as at.%.

[0087] When the input text is "CoCrMnNiMo0.2", the system recognizes the value followed by the numerical subscript 0.2 after Mo and calculates according to the same formula. If the sample name contains multiple variable values, the system can generate multiple coating entries based on the variable values.

[0088] In this way, the system can convert common variable subscript chemical formulas in materials literature into comparable atomic percentage data.

[0089] Example 6: Intelligent Extraction Intelligent component extraction using a large language model includes the following operations: Construct prompt information based on the text to be extracted, the sample name, and / or the image context; The instruction is to output the coating representation, element list, content unit, extraction mode, explicit component information, subscript information, and confidence level in a pre-defined JSON structure using the large language model. Perform JSON parsing, field validation, element symbol validation, and component calculation on the content returned by the large language model; When the large language model returns equimolar components or variable subscript components, the element content is calculated.

[0090] For example, when the input text is "The coating was prepared using cobalt, chromium, iron, and nickel in equal atomic proportions.", rule extraction might not be able to directly obtain the chemical formula from the full English names of the elements. The intelligent extraction module is invoked, prompting the large language model to identify the coating or alloy composition and requesting a JSON structure. After the model returns Co, Cr, Fe, Ni, and the equimolar pattern, the system locally calculates each element as 25 at.%.

[0091] When the input text is "coating composition is 22% cobalt, 18% chromium, balance iron", the intelligent extraction module can map the Chinese element names to the element symbols Co, Cr, and Fe, and identify iron as the balance. The system then performs balance calculation and verification.

[0092] To reduce the illusion of large language models, the system does not directly accept conclusions from free text. Instead, it requires the model to return preset fields and performs secondary verification on element symbols, extraction patterns, units, and the sum of components.

[0093] Example 7: Bidirectional conversion between mass percentage and atomic percentage includes the following operations: When converting from mass percentage to atomic percentage, it is calculated according to the formula at_i=(wt_i / M_i) / Σ(wt_j / M_j)×100, where wt_i is the mass percentage of the i-th element and M_i is the atomic weight of the i-th element; When converting from atomic percentage to mass percentage, it is calculated according to the formula wt_i=(at_i×M_i) / Σ(at_j×M_j)×100, where at_i is the atomic percentage of the i-th element; After conversion, the element content will be formatted to a uniform precision, and the content unit will be updated.

[0094] For example, when a user requests to convert wt.% to at.%, the system reads the mass percentage and atomic weight of each element. For element i, its molar contribution is calculated as wt_i / M_i, and then the molar contributions of all elements are summed and normalized to obtain at_i=(wt_i / M_i) / Σ(wt_j / M_j)×100.

[0095] When a user requests to convert at.% to wt.%, the system calculates the quality contribution of each element as at_i×M_i, and then normalizes it to obtain wt_i=(at_i×M_i) / Σ(at_j×M_j)×100.

[0096] Taking a typical input as an example: CoCrFeNiMox (x=0.2, molar ratio); CoCrMnNiMo0.2; When outputting the results, you can choose the unit yourself (e.g., ...). Figure 7 (The part highlighted in red in the middle); Selecting different units will allow the system to automatically calculate the conversion. Clicking "Preview Results" will show the conversion results, as shown below. Figure 8 As shown.

[0097] After conversion, the system will uniformly retain the results to a preset number of decimal places and perform a more stringent threshold check on the converted sum. For example, the total of the converted results should be close to 100; if the deviation exceeds the threshold due to rounding or abnormal data, the system will output a warning.

[0098] Example 8: Batch management and export include the following operations: Assign incremental batch numbers to one or more extraction or conversion operations; Record the creation time and coating quantity for each batch; Summarize the element symbols appearing in all batches as a dynamic header; Write the batch number, coating expression, extraction mode, extraction status, status description, content of each element and content unit into an electronic spreadsheet file.

[0099] Specifically, the system generates an incrementing batch number for each save operation and records the batch's creation time. A batch can contain one or more coating composition data entries. Users can view the batch number, time, and coating quantity of all batches, and can also clear batch data.

[0100] During export, the system iterates through all batches and coating data, collecting all element symbols that have appeared and dynamically generating element columns. Each row records a coating entry, including batch number, coating expression, extraction mode, extraction status, status description, element content, and unit. The exported spreadsheet filename includes a timestamp, making it easy for users to distinguish between different export results.

[0101] This invention can not only process single text input, but also incorporate multi-source heterogeneous material literature data such as text input, image input, and sample names into a unified "rule-first, intelligent supplementation" extraction system after preprocessing and contextualization. This solves the technical problem of the difficulty in automatically parsing material literature data into traceable and verifiable structured coating composition data due to the diverse sources and inconsistent expression forms. Furthermore, this invention demonstrates the technical solution of constraining OCR recognition results and intelligent recognition results with rule extraction and local calculation.

[0102] To verify the applicability and effectiveness of the multimodal intelligent composition extraction and conversion system for materials science described in this invention in practical scientific research data processing scenarios, this system was used to process and verify all relevant literature data collected by the applicant in the field of materials science from 2023 to 2025. The verification objects mainly included alloy composition, element content, composition units, sample numbers, and related data description information involved in materials literature. The data sources included natural language descriptions in the main text of the literature, standardized or non-standard alloy composition expressions, tabular composition data, and some image-based composition information.

[0103] During the verification process, the literature excerpts, composition expressions, or composition table images to be processed are first input into the system. The system then automatically calls the corresponding composition extraction module based on the specific format of the input data. For composition expressions with relatively standardized formats, the system uses a rule engine to complete element identification, content analysis, and unit determination. For natural language descriptions, variable alloy compositions, complex tables, or image-based data, the system uses an intelligent recognition module and an optical character recognition module to complete content analysis and convert the extraction results into structured composition data. For cases where both atomic percentage and mass percentage expressions exist in the literature, the system further calls the unit conversion module to standardize the atomic percentage and mass percentage, and verifies the element types, element contents, unit types, and sample information in the result preview interface. After verification, the batch management module is used to save data from different literature sources, different samples, or different coating systems in batches, and finally export them as standardized Excel data files.

[0104] Based on the above verification, this system can cover various compositional expression formats commonly used in materials literature data processing, including equimolar alloy expressions, explicit percentage composition expressions, alloy composition expressions with variables, natural language composition descriptions, and tabular composition data. With the assistance of this system, the applicant completed the systematic collection and standardization of all materials literature data related to high-entropy alloys from 2023 to 2025, ultimately generating 675 high-quality structured data entries. All data entries contain core compositional information that can be used for subsequent scientific research analysis, meeting the basic data needs for materials database construction, machine learning modeling, composition screening, and subsequent experimental design.

[0105] Compared with traditional manual copying, conversion, and table entry methods, this system demonstrates the following technical advantages in practical use: First, it can transform component information scattered across the main text, tables, and images of documents into structured data, reducing the workload of repeatedly identifying and entering data between different document formats. Second, it can uniformly process different component units, reducing data inconsistencies caused by mixing atomic percentages and mass percentages. Third, it can achieve continuous data processing, batch saving, and unified summarization through batch management and Excel export functions, improving the standardization and traceability of the materials literature data processing process. Fourth, it can adapt to the more complex alloy composition expression methods in the field of materials science, especially high-entropy alloys, near-equimolar alloys, variable composition systems, and balance element descriptions, thereby improving the system's practicality in materials composition data processing tasks.

[0106] The above verification results demonstrate that the system described in this invention can extract, convert, manage, and export multi-source, multi-format composition information in practical materials literature data processing, effectively assisting in the standardized processing of large volumes of materials literature data. This system not only improves the efficiency and consistency of materials composition data processing but also provides a structured data foundation for subsequent materials database construction, machine learning model training, alloy composition design, and experimental scheme optimization, exhibiting significant engineering application and research support value.

[0107] The embodiments given above are preferred examples for implementing the present invention, and the present invention is not limited to the above embodiments. Any non-essential additions or substitutions made by those skilled in the art based on the technical features of the present invention are within the protection scope of the present invention.

Claims

1. A method for extracting and converting coating composition information from materials literature, characterized in that, include: S1. The input processing module acquires the literature data of the material to be processed, which includes text data, image data and sample name; S2. When the literature data to be processed includes image data, the image recognition module performs format verification, image preprocessing, and optical character recognition on the image data to obtain the recognized text. S3. The text data and the identified text are normalized to obtain the text to be extracted; and the coating representation and candidate element symbols are extracted from the text to be extracted and the sample name. S4. The rule extraction module sequentially performs phase coating extraction, explicit component extraction, equimolar component extraction, and variable subscript component extraction on the text to be extracted. The extraction of the added phase coating involves detecting whether it is an added phase coating and identifying the matrix alloy, the chemical formula of the added phase, the content of the added phase, and the unit. Clearly define the components: identify the correspondence between element symbols and numerical content, and make up for the remaining amount to ensure that the content of each element is logically closed. Equimolar extraction: After detecting keywords including equimolar / equiatomic ratio, the effective element set is extracted from the continuous element markers, and the content of each element is normalized to a percentage under the same benchmark. Variable subscript extraction: Parse the subscript variables and their values, and normalize the content of each element into a percentage under a unified standard according to a predetermined proportional relationship; If the rule extraction module does not obtain effective components, or is forced to extract intelligent components, the intelligent extraction module calls the large language model to extract intelligent components from the text to be extracted. S5. Generate coating component data based on the component types extracted by the rule extraction module or intelligent extraction module; the coating component data includes coating expression, element symbols, element content, content unit, extraction mode, extraction status and status description; S6. The component conversion module performs bidirectional conversion between mass percentage and atomic percentage of the coating component data based on the element atomic weight data. S7. The data verification module performs element validity verification, total component verification, and single element abnormality percentage verification on the coating component data. S8, the batch management module saves the verified coating composition data by batch; the file export module generates exportable structured files and exports them.

2. The method for extracting and converting coating composition information from material literature as described in claim 1, characterized in that, The image preprocessing and optical character recognition include: Verify image format and file size; Convert the image to a uniform color mode; When the image size is below a preset threshold, the image is magnified. The recognized text is obtained using an optical character recognition engine that supports mixed Chinese and English recognition; The identified text is then processed to perform whitespace character normalization and common recognition error correction. The normalization of the text data and the identified text includes: First, perform basic normalization, including removing leading and trailing whitespace, to form the string to be parsed; Then, based on the alloy / coating marking pattern in the sample name or text, extract or complete the coating marking information.

3. The method for extracting and converting coating composition information from material literature as described in claim 1, characterized in that, The specific extraction of the defined components includes the following operations: Use pattern matching to extract element symbols and content, and standardize the element symbols into a standard chemical nomenclature format; Identify at least one unit expression among mass percentage, atomic percentage, mass fraction, atomic fraction, and their English expressions, and standardize the units to a unified standard; Identify surplus element keywords, wherein the surplus element keywords include at least one of Bal., balance, surplus, remainder, or surplus; When there is a surplus element and the total known element content is less than 100, the surplus value is obtained by subtracting the total known element content from 100, and the surplus value is allocated to the surplus element.

4. The method for extracting and converting coating composition information from material literature as described in claim 1, characterized in that, The extraction of equimolar components specifically includes the following operations: The text to be extracted was tested for keywords including equimolar, equiatomic, equal molar, equal atomic ratio, and equiatomic. Multiple element symbols are parsed from the coating representation or the text to be extracted; The atomic percentage of each element is calculated by dividing 100 by the number of elements, and the content unit is set to atomic percentage.

5. The method for extracting and converting coating composition information from material literature as described in claim 1, characterized in that, The extraction of variable subscript components specifically includes the following operations: Identify coating representations containing variable subscripts, coating representations containing numerical subscripts, or coating representations containing underlined numerical values; Determine matrix elements, variable elements, and variable values ​​from the coating representation, the text to be extracted, and / or the sample name; When the variable value is x and the number of matrix elements is n, the atomic percentage of each matrix element is calculated as 100 / (n+x), and the atomic percentage of the variable element is calculated as x×100 / (n+x); When the same variable subscript component contains multiple variable values, an independent coating component data is generated for each variable value, and the matrix element, variable element, variable value and subscript mode type are recorded in the coating component data.

6. The method for extracting and converting coating composition information from material literature as described in claim 1, characterized in that, Intelligent component extraction using a large language model includes the following operations: Construct prompt information based on the text to be extracted, the sample name, and / or the image context; The instruction is to output the coating representation, element list, content unit, extraction mode, explicit component information, subscript information, and confidence level in a pre-defined JSON structure using the large language model. Perform JSON parsing, field validation, element symbol validation, and component calculation on the content returned by the large language model; When the large language model returns equimolar components or variable subscript components, the element content is calculated.

7. The method for extracting and converting coating composition information from material literature as described in claim 1, wherein the bidirectional conversion between mass percentage and atomic percentage includes the following operations: When converting from mass percentage to atomic percentage, it is calculated according to the formula at_i=(wt_i / M_i) / Σ(wt_j / M_j)×100, where wt_i is the mass percentage of the i-th element and M_i is the atomic weight of the i-th element; When converting from atomic percentage to mass percentage, it is calculated according to the formula wt_i=(at_i×M_i) / Σ(at_j×M_j)×100, where at_i is the atomic percentage of the i-th element; After conversion, the element content will be formatted to a uniform precision, and the content unit will be updated.

8. The method for extracting and converting coating composition information from material literature as described in claim 1, characterized in that, The element validity verification, total component verification, and single element abnormality percentage verification include the following operations: Match element symbols with a preset element table; Determine whether the sum of the extracted component data falls within the first threshold range; Determine whether the sum of the components obtained from the conversion is within the second threshold range; A warning is generated when the content of a single element exceeds a preset threshold for that single element. The error messages, warning messages, and verification details are output in association with the coating composition data.

9. The method for extracting and converting coating composition information from material literature as described in claim 1, characterized in that, The process of saving in batches and generating exportable structured files includes the following operations: Assign incremental batch numbers to one or more extraction or conversion operations; Record the creation time and coating quantity for each batch; Summarize the element symbols appearing in all batches as a dynamic header; Write the batch number, coating expression, extraction mode, extraction status, status description, content of each element and content unit into an electronic spreadsheet file.

10. A system for extracting and converting coating composition information from materials literature, characterized in that, include: The input processing module acquires text data, image data, and / or sample names; The image recognition module performs format verification, preprocessing, and optical character recognition on image data. The rule extraction module performs phase coating detection, explicit component extraction, equimolar component extraction, and variable subscript component extraction on the recognized text. The extraction of the added phase coating involves detecting whether it is an added phase coating and identifying the matrix alloy, the chemical formula of the added phase, the content of the added phase, and the unit. Clearly define the components: identify the correspondence between element symbols and numerical content, and make up for the remaining amount to ensure that the content of each element is logically closed. Equimolar extraction: After detecting keywords including equimolar / equiatomic ratio, the effective element set is extracted from the continuous element markers, and the content of each element is normalized to a percentage under the same benchmark. Variable subscript extraction: Parse the subscript variables and their values, and normalize the content of each element into a percentage under a unified standard according to a predetermined proportional relationship; The intelligent extraction module calls a large language model to perform intelligent component extraction on the text to be extracted when the rule extraction fails to obtain effective components or when forced intelligent extraction is specified. The composition conversion module performs bidirectional conversion between mass percentage and atomic percentage based on element atomic weight data. The data verification module performs element validity verification, component sum verification, and single element abnormality percentage verification. The batch management module saves coating composition data for different batches; The file export module generates a structured file containing batch information, extraction status, and element content.

Images