One Domain, Four Schemas

TL;DR for operators

Scientific-literature extraction is not only a question of whether an AI system can find the right facts. It also needs a representation capable of preserving the distinctions engineers will later use.

Sadruddin et al. build four separate machine-validatable process schemas for atomic layer deposition (ALD) and atomic layer etching (ALE), divided again between experiments and simulations.¹ The split is consequential: the schemas contain between 167 and 275 properties, while required-property counts range from just 3 in ALD simulation to 56 in ALE simulation.

They nevertheless share a semantic core. Six physical quantity kinds—Energy, Length, Pressure, Temperature, Thickness, and Time—appear across all four schemas, with quantitative fields grounded in standardized quantity and unit semantics.

The practical result is more than cleaner storage. In the paper’s ALD extraction demonstration, the schema becomes part of the control layer around an LLM: extracted records can be checked for required structures, datatypes, permitted units, quantity kinds, and defined numerical ranges before entering downstream systems.

For semiconductor R&D, the design implication is clear: normalize what is genuinely common, but preserve process- and perspective-specific structure where the science demands it. The paper demonstrates this operationally for ALD experimental literature; equivalent extraction performance across the other three schemas remains unestablished.

Structured data fails when the structure is too generic

A semiconductor data team trying to convert hundreds of process papers into searchable records faces an apparent normalization problem. Authors use different terminology, report values in different units, scatter conditions across prose and tables, and omit fields that another paper treats as central.

The tempting response is to design one sufficiently comprehensive template and make every paper fit it.

The paper shows why that can be the wrong abstraction. Its four schemas differ not just in vocabulary but in what information they require and how deeply that information is organized.

These differences reflect different scientific tasks. Experimental representations emphasize operating conditions and measured material or device properties. Simulation representations allocate more structure to computational methods, surface models, reaction mechanisms, and predicted quantities.

That changes the extraction problem. An AI system can correctly identify a temperature, reactant, or material and still produce a poor scientific record if the surrounding representation omits the relationships and constraints needed for that kind of process.

Four schemas preserve distinctions that normalization cannot remove

Each schema instance represents one literature-reported process, not an entire publication. A single paper can therefore yield multiple process records.

The schemas use JSON Schema to specify what a valid record looks like: which fields can appear, which structures are required, and what kinds of values those fields may contain. This creates a closed-world validation layer rather than relying entirely on whatever structure an extraction model happens to produce.

The four designs still retain common concepts. Pressure appears in every schema. Reactants or precursors, process conditions, material outcomes, and gas or reactant flows also recur, although sometimes under different structures and at different granularities.

Five manually normalized domain properties are common to all four: film chemical composition, pressure, the primary precursor or reactant, reactant flow rate, and substrate material.

The important point is not that the schemas are unrelated. It is that a common vocabulary does not imply a common information architecture.

A simulation record and an experimental record can refer to the same physical process while requiring different evidence to be useful. Collapsing both into one universal structure risks either discarding specialized information or producing a template so permissive that structural validation becomes weak.

Shared quantities provide the interoperability layer

Numbers create a second problem. Recording that a process operated at “100” is useless without knowing whether the value refers to temperature, pressure, time, or something else. Even a labeled quantity remains difficult to compare if equivalent values use incompatible unit conventions.

The authors address this by grounding quantitative properties in QUDT, an ontology for quantities, units, dimensions, and types. Each grounded property can explicitly represent its magnitude, physical quantity kind, and permitted units.

Across all four schemas, six quantity kinds survive as a common core:

Energy, Length, Pressure, Temperature, Thickness, and Time.

This is where ontology grounding becomes operational rather than terminological. Two records can use different surface labels or units while still identifying the same underlying physical quantity.

The design therefore separates two kinds of harmonization. Shared physical meaning can be standardized across the domain, while process-specific structure remains distinct.

For scientific-data systems, that is a more useful target than making every record look identical.

The schema can become an AI validation layer

The paper’s most direct operational demonstration uses the ALD experimental schema as a blueprint for extracting structured information from ZnO and IGZO ALD publications drawn from AtomicLimits.

An agentic LLM workflow extracts process information, normalizes chemical entities to PubChem identifiers, and produces records that are checked against the schema.

The important mechanism is the validation step. The schema can reject or flag outputs that violate expected datatypes, omit required structures, use disallowed units, mismatch quantity kinds, or exceed defined numerical ranges.

That changes the role of the schema. It is no longer merely documentation for a database designer. It becomes part of the runtime quality-control system around extraction.

The same corpus-level extraction also reveals something about the literature itself: chemistry-related fields such as deposited materials, precursors, co-reactants, and process gases were populated relatively consistently, while growth-per-cycle, dosing-duration, and purge-duration information appeared less uniformly.

Structured extraction can therefore expose not only model failures but also systematic gaps in scientific reporting.

For semiconductor data teams, representation becomes part of pipeline design

The paper directly establishes four curated resources, their structural differences, their shared quantitative semantics, and one schema-guided extraction use case.

A broader operational implication follows from those pieces.

A semiconductor R&D organization building literature search, process analytics, knowledge graphs, or AI-assisted process-development tools could place schema validation between extraction and downstream ingestion:

paper → extraction → schema validation → normalized process record → analytics / knowledge graph / AI

That placement matters. Without it, an extraction system may pass fluent but structurally incomplete records directly into systems that assume the data are comparable.

The potential value is therefore not simply less manual data entry. It is a stronger contract between literature extraction and downstream computation: process records have explicit structure, quantitative semantics, and at least some machine-checkable physical constraints before other systems depend on them.

The four-schema design also suggests an architectural principle for scientific AI systems: reuse semantic components where the domain genuinely overlaps, but allow separate record structures when experimental perspective or process mechanics change what information is decision-relevant.

The extraction evidence currently stops at ALD experiments

The practical demonstration should not be read as validation of a complete ALD/ALE literature-mining platform.

The extraction use case applies the ALD experimental schema. The paper does not establish equivalent extraction performance for ALD simulations, ALE experiments, or ALE simulations.

Ontology alignment is also concentrated on quantitative properties. Materials, processing methods, structures, characterization activities, and simulation activities do not yet receive the same breadth of semantic grounding.

Finally, the authors leave wider corpus evaluation and further domain-expert refinement as future work. The paper mainly evaluates the resulting representations through literature-guided construction, structural comparison, and a methodological demonstration rather than through causal or standardized statistical performance estimates.

Those boundaries still leave a concrete design result: extraction quality depends partly on choosing the right target representation before the model starts filling fields.

Representation is part of extraction

Scientific AI systems are often evaluated on whether they can retrieve the right fact from an article. This work points to a second question: what structure will receive that fact, and what constraints will determine whether the resulting record is acceptable?

For ALD and ALE, one universal schema would erase differences the authors found worth preserving. Four specialized schemas, connected through shared quantity semantics, provide a more disciplined compromise.

That makes the central contribution less about producing another metadata format and more about where representation sits in the workflow. When the schema specifies both what the AI should extract and what the system will accept, data modeling becomes part of the extraction mechanism itself.

Cognaptus: Automate the Present, Incubate the Future.