Can AI Mark Physics, Chemistry and Maths? How GradeDrive Handles Complex STEM Questions

The Problem That Stops Most AI Marking Tools in STEM Subjects

Ask most AI marking tools to mark a History essay or an English Language response, and they will do a reasonable job. Ask them to mark a GCSE Physics calculation, a Chemistry paper with structural formulae, or an A level Maths question with five lines of working, and the results are much less reliable.

This is not a minor gap. In UK secondary schools, STEM subjects account for a significant proportion of assessed work. Physics, Chemistry, Biology, and Mathematics are among the most marked subjects in the country. A marking tool that cannot handle them is not a general solution for teacher workload — it is a partial one.

The reason most tools struggle with STEM is not the AI model itself. It is the reading and extraction step that happens before any AI can assess a response. STEM exam papers contain content that does not look like text, does not behave like text, and cannot be reliably interpreted without understanding the structure behind it.

GradeDrive's approach to this problem is what makes it usable for STEM teachers. This post explains what that approach involves and what it delivers in practice.

Why STEM Responses Are Structurally Different

A GCSE or A level STEM response can contain several types of content simultaneously — in the same answer box, on the same line, sometimes overlapping.

Mathematical workings are written non-linearly. A student solving a multi-step Physics calculation might write the formula on one line, substitute values on the next, write an intermediate result partway down the page, and circle the final answer in a different location. The logical sequence of the working is clear to a human reader because they understand the mathematical structure. It is not clear from the spatial layout of marks on a page.

Chemical equations use a notation system — element symbols, subscripts, superscripts, arrows, state symbols, structural bonds — that is fundamentally different from prose text. A student writing a balanced equation for the combustion of methane is not writing words; they are writing a structured symbolic representation that obeys specific rules. Reading it correctly requires understanding those rules, not just recognising characters.

Structural formulae go further. A student drawing the structural formula for ethanol by hand is producing a two-dimensional diagram with bonds represented as lines, atoms as letters, and spatial arrangement carrying chemical meaning. No general-purpose OCR system reads this correctly, because general-purpose OCR was designed for text documents, not structural chemistry.

Diagrams and labelled illustrations appear throughout GCSE Biology, Physics, and Geography papers. A student drawing a neurone, a circuit, a plant cell, or a geological cross-section and labelling the parts is producing a response where the labels and the diagram must be read together to be meaningful. The label "axon terminal" attached to the correct part of a neurone diagram means something different from the same label attached to the wrong part — and an AI marking system that cannot tell the difference will make systematic errors on diagram questions.

GradeDrive's Extraction and Structuring Pipeline

GradeDrive handles STEM content through a purpose-built extraction pipeline that processes different content types differently before any AI assessment takes place.

The pipeline does not treat a STEM exam response as a page of text to be read sequentially. It identifies content regions — prose, numerical workings, equations, diagrams, labels — and applies different extraction methods to each. The outputs of those extractions are then structured into a representation of the student's response that the AI can assess against the mark scheme.

This is the "hybrid AI and non-AI" approach that distinguishes GradeDrive from tools that route everything through a single large language model. Some parts of the extraction process — particularly the identification and interpretation of mathematical structure and chemical notation — are handled by specialised non-AI components that apply the rules of the notation system directly, without relying on a language model to infer them. The AI components then operate on the structured output, not the raw handwritten image.

The practical effect is significantly more reliable extraction of STEM content than a pure-LLM approach produces. A language model asked to interpret a photograph of handwritten chemistry working will produce a plausible-looking output that may or may not be correct. A structured extraction pipeline that identifies the equation, parses the chemical notation according to its formal rules, and outputs a validated symbolic representation produces a reliably correct one.

Mathematics: Workings, Method Marks, and Error Carried Forward

Mathematics marking involves a set of conventions that are specific to the subject and largely invisible to non-specialists.

Method marks are awarded for the correct application of a mathematical process, regardless of whether the final numerical answer is correct. A student who sets up a quadratic equation correctly, attempts to solve it using the quadratic formula, but makes an arithmetic error in the final step can still earn the method marks for the setup and the approach. Recognising whether the method is correct requires reading the workings, not just the answer.

Error carried forward (ECF) means that if a student makes an error early in a multi-part question and then uses their incorrect value correctly in subsequent parts, they can still earn marks for those subsequent parts. Applying ECF correctly requires tracing the student's logic through the whole question — understanding that the value on line 4 came from the incorrect answer on line 2, and that it was used consistently and correctly from that point onwards.

Shown workings are required or encouraged on almost every calculation question. The mark scheme awards marks for specific steps in the working, not just the final answer. A student who writes the correct answer with no working shown typically earns fewer marks than one who shows a complete, correct method.

GradeDrive reads mathematical workings in sequence, identifies the steps the student has taken, and assesses them against the mark scheme's method mark structure. ECF logic is applied where the scheme specifies it. The teacher review step is particularly valuable on calculation questions where a student's working is unconventional but correct — the AI flags these for human check rather than making a potentially incorrect automatic decision.

Chemistry: Equations, Formulae, and Structural Drawings

Chemistry presents two distinct extraction challenges: symbolic equations and structural representations.

Symbolic equations — including word equations, balanced symbol equations, ionic equations, and half equations — use a notation that must be read correctly for the mark scheme comparison to work. A student writing "2H₂ + O₂ → 2H₂O" has written a balanced equation with specific requirements: correct formulae, correct coefficients, correct arrow type, and correct state symbols where required. Each of these elements corresponds to a potential mark or deduction in the scheme.

GradeDrive's chemistry extraction parses the components of handwritten equations — identifying element symbols (distinguishing capital and lowercase letters that change the meaning, such as Co vs CO), reading subscripts and superscripts in context, and validating the balance and structure of the equation against the mark scheme requirements.

Structural formulae require a different approach. A student drawing the displayed formula of an organic compound is producing a spatial diagram where bond angles, atom positions, and connectivity all carry chemical meaning. GradeDrive's extraction pipeline identifies structural drawings, reads the atom labels and bond connections, and constructs a representation of the structure that can be compared to the mark scheme's expected answer — including partially correct structures where some features are right and others are wrong.

Biology and Physics: Diagrams with Labels

Diagram questions appear across Biology and Physics at both GCSE and A level. A student might be asked to draw and label a cell, annotate a force diagram, complete a circuit diagram, or add labels to an anatomical illustration.

The marking challenge is not just reading the labels — it is determining whether the labels are in the right place. A label that says "nucleus" is only correct if it points to the nucleus. A label that says "cell membrane" pointing to the cytoplasm is incorrect, even though the label text is accurate.

GradeDrive handles labelled diagram questions by processing the spatial relationship between label text and the diagram regions they point to, not just the text content of the labels. This allows the marking system to distinguish between a correct label correctly placed and a correct label incorrectly placed — a distinction that matters for the mark.

What This Means for Subject Teachers

For Physics, Chemistry, Biology, and Maths teachers, the practical implication is this: GradeDrive can mark the papers you actually set, using the responses your students actually produce.

You do not need to avoid calculation questions because the AI can't read the workings. You do not need to exclude Chemistry papers because of equation notation. You do not need to mark diagram questions by hand while letting the AI handle only the prose questions. The full paper — all question types, all subjects — goes through the same upload-and-review workflow.

The review step remains valuable for the edge cases: the student whose working is unusual but valid, the structural formula that is drawn in a non-standard orientation, the diagram where a label is ambiguous. GradeDrive flags these for teacher attention rather than making an automatic call. But for the majority of responses — the clear correct answers, the clearly incorrect ones, and the responses that earn partial credit according to the scheme — the AI handles the marking and the teacher confirms it.

Frequently Asked Questions

Can GradeDrive mark GCSE Physics papers? Yes. Physics papers including calculation questions with shown workings, diagram questions, and extended writing are all supported. GradeDrive reads mathematical workings in context and applies method mark and ECF logic.

Can GradeDrive read chemical equations written by hand? Yes. GradeDrive's extraction pipeline parses handwritten chemical equations — including balanced symbol equations, ionic equations, and half equations — reading element symbols, subscripts, superscripts, and state symbols correctly.

Can GradeDrive interpret structural formula drawings? Yes. Handwritten structural formulae — including displayed formulae for organic compounds — are processed by GradeDrive's structural extraction component, which reads atom labels and bond connectivity rather than just the text content of the response.

Can GradeDrive mark A level Maths papers? Yes. A level Maths is supported, including the extended calculation chains and proof-based questions typical of A level. GradeDrive reads working sequences and applies method mark logic. The teacher review step is particularly useful on unconventional-but-correct workings.

What happens with diagram labelling questions? GradeDrive processes the spatial relationship between label text and diagram regions, not just the label text. This allows it to mark labelled diagrams correctly — distinguishing between a correctly placed label and a correctly worded label in the wrong position.

Does GradeDrive work for all STEM subjects at A level? Biology, Chemistry, Physics, and Maths at A level are all supported. Geography and Psychology at A level are also supported for the extended writing and shorter response components.

Try GradeDrive free — upload a STEM paper and see how the extraction and marking handles your subject.