Quick answer: Making an established analysis run faster — scripting model set-up, parallelising a solver, automating result extraction — is generally not a core R&D activity: the analysis method and the technical result it produces are already known, and what is uncertain is effort. What may be core is developing a surrogate, discretisation or optimisation method whose achievable accuracy, convergence or solution quality against validated reference results could not be determined in advance and could only be determined by a systematic progression of work conducted to generate new knowledge. Automation built solely to execute or measure a genuine core experiment may separately require assessment as a supporting activity, including the dominant-purpose test where applicable. Analysis undertaken to demonstrate compliance with a statutory requirement or standard is excluded from being a core R&D activity under s 355-25(2). You self-assess.
7 August 2026 — this article describes the current rules. The 2026–27 Federal Budget announced proposed R&DTI reforms for income years starting on or after 1 July 2028. Until any amendments take effect, the R&DTI continues to be administered under the current legislation.
The same automation project recurs across engineering consultancies and design-software teams: an analysis that takes a senior engineer three days is rebuilt to run in three minutes. Parametric geometry regenerates the model. Scripts mesh, solve and post-process without a human in the loop. An optimisation loop sweeps the design space overnight. Increasingly, a machine-learning model trained on past solver runs predicts the answer instead of computing it.
The question this article answers is narrower than whether that work is valuable: which parts of it, if any, may be eligible R&D activities under the R&D Tax Incentive (R&DTI), and which parts are skilled engineering that is not R&D. I write as a Registered Research Service Provider (RSP000047) at Lot Fourteen in Adelaide. We supply research capability and assist with planning and conducting technical R&D activities within our registered research fields, we do not give tax advice, and the rules below belong to the ATO, AusIndustry and the legislation, not to us.
Scope note: this piece is about the analysis and simulation layer. Where the analysis exists to demonstrate that a design conforms to the National Construction Code or an Australian Standard, our property and construction coverage deals with compliance testing versus developing the system being tested.
Speed Is an Effort Problem, Not a Knowledge Problem
A core R&D activity has to satisfy both limbs: the outcome could not be known or determined in advance on the basis of current knowledge, information or experience, and could only be determined by applying a systematic progression of work that is based on principles of established science and proceeds from hypothesis to experiment, observation and evaluation and leads to logical conclusions; and the activity is conducted for the purpose of generating new knowledge (business.gov.au).
The established-science requirement inside that progression does real work in this discipline: a surrogate programme that is a run of architecture and hyperparameter tweaks, kept because one of them scored best, is a search rather than a progression based on principles of established science — what grounds the work is the continuum mechanics, heat transfer or numerical analysis that tells you which features, error measures and regime boundaries the trials should be probing.
Where a validated finite-element or thermal model already exists and a competent professional can determine in advance how the automated workflow will reproduce that established analysis, the automation is generally implementation rather than core R&D. Automating it changes who produces the answer and how long it takes. No new knowledge about the physics, the method or its limits is generated; labour is removed. The uncertainty that remains is scheduling and integration risk — whether the geometry kernel behaves, whether the licence server copes, whether the mesh script survives edge cases — which is business and engineering risk rather than the statutory unknown.
Our software and AI material and our what does not qualify page cover the general version of that line. Below is the version specific to analysis and simulation.
The question that separates the two:
At the outset, could a competent professional in this field have determined, from existing knowledge, information or experience, what accuracy (or convergence, or stability) this method would achieve on this class of problem — or could that only be established by a systematic progression of experimental work?
Where the team knew the method would work and the remaining task was construction, the technical boundary falls outside experimentation. Where the achievable technical result could not be determined and was established by experiment against validated reference solutions, there is a delineable technical work package — which the company and its registered tax agent then assess against every statutory requirement and exclusion.
Where the Analysis Layer Usually Sits
Usually effort, not statutory uncertainty
May contain a core R&D activity
Scripting model set-up, meshing, batch submission and result extraction for an established solver
Developing a surrogate or reduced-order model whose achievable accuracy against validated reference solutions could not be predicted for that class of geometry or loading
Parallelising or porting an existing analysis to more hardware to shorten run time
Developing a numerical or discretisation scheme where whether it converges — and to what error — is genuinely unknown
Parametric geometry that regenerates a model from a spreadsheet of inputs
Establishing whether an optimisation formulation can reliably reach a defined solution quality on a design space where existing methods have no predictive basis
Applying a standard optimisation package to a standard problem with expected behaviour
Establishing the domain of validity and error bounds of a learned model where behaviour outside the training distribution is the unknown being investigated
Building a user interface, dashboard, licensing or job queue around the analysis
Developing an active-learning or adaptive-sampling strategy where the achievable accuracy-per-solver-run trade-off is unknown
Two guardrails on the right-hand column:
Using machine learning does not make it R&D: AusIndustry's AI sub-guide to the software development sector guide states that applying existing AI techniques in a routine way is not, of itself, an eligible activity. Training a network on 50,000 solver runs inside a well-understood parameter range, to an accuracy a competent professional could have anticipated, is a build.
"Faster" is not a hypothesis: "Reduce analysis time by 90%" is a project objective. A hypothesis is a testable proposition about an unknown outcome — that a particular surrogate formulation holds predicted peak stress within a stated error band, across a stated geometry envelope, against validated reference solutions — with the failure condition fixed before the trials begin. Where elapsed time is the only quantity being measured, the activity is almost certainly not core R&D.
The Standards-Compliance Trap
Subsection 355-25(2) of the Income Tax Assessment Act 1997 excludes several categories from being core R&D activities, including activities associated with complying with statutory requirements or standards (business.gov.au — excluded R&D activities).
For an analysis team that reaches further than people expect. Where the purpose of the analysis is to demonstrate compliance with a requirement or standard imposed by legislation — for example, a verification or certification activity required under the applicable regulatory framework — the statutory-compliance exclusion may apply. Conducting R&D in a regulated environment is not itself excluded merely because the resulting design must ultimately comply with legal requirements.
The exclusion bars those activities from being core; it does not put them outside the program. An activity that is of an excluded kind, produces goods or services, or is directly related to producing goods or services qualifies as a supporting R&D activity only if it is undertaken for the dominant purpose of supporting a core R&D activity (business.gov.au). Where there is no core activity, there is nothing for supporting activities to support.
Experimental Design and Records That May Strengthen the Evidence
Depending on the technical question, the following practices may help demonstrate a systematic progression of work and distinguish experimental activities from surrounding implementation:
Define and preserve the benchmark data where appropriate: Freezing or versioning reference data can help ensure results are compared on a consistent basis. Reference solver runs, input decks, solver version and tolerances are pinned and recorded. A corpus that grows during the year cannot support a before-and-after error claim, because the denominator moved.
Mesh and solver tolerance are controlled, not incidental: A convergence study fixes the mesh density and the residual tolerance the reference runs use, so that surrogate error is separated from discretisation error. Without it, an observed error band cannot be attributed to the method under test.
Tuning is separated from evaluation: Hyperparameter search and formulation changes run against a validation split; the holdout is opened once, at the end, against the pre-registered metric. Repeated use of the same holdout during tuning may weaken its value as independent evaluation evidence, so the evaluation design and any changes should be documented.
Leakage is designed out by geometry family, not by row: For highly related parametric data, a random row-level split may create leakage or overstate generalisation; the split method should be technically justified for the question being tested. Splitting by build-up, section family or load case is what makes the extrapolation question answerable.
Model configurations are versioned like code: Architecture, feature set, training seed, data split manifest and the resulting metric are recorded per version, including the versions that were abandoned. Failed formulations are the clearest contemporaneous evidence that the result could not be known in advance.
Engineering time is coded at entry into three streams: Experimentation, productisation of the tool, and client delivery. Reconstructing that split from a project ledger at registration is the single most common failure mode in this discipline, because a simulation engineer moves between all three inside one day.
A Worked Example — Hypothetical and Illustrative Only
The following scenario is entirely hypothetical. It illustrates where a technical boundary falls; it is not a statement that any of this work would be eligible, and eligibility would be self-assessed by the company against the statutory tests.
An Adelaide engineering consultancy wants a fast predictor of transient thermal behaviour in ventilated façade assemblies, to replace a three-day solver run during early design.
Envelope and baseline: The team fixes the geometry envelope: four cavity build-ups (cavity depth 25–60 mm), three cladding materials, insulation thickness 40–120 mm, four orientations, with one summer and one winter diurnal temperature and solar-flux profile held constant across every trial. The existing screening correlation predicts peak interior surface temperature to about ±2.1 °C on that envelope. The target is set at ±0.5 °C at the 95th percentile of the holdout, because below that band the surrogate would not change a design decision relative to the full solver. Failure is defined in advance as not reaching that band after the planned trial sequence.
Reference data: 3,200 transient solver runs are generated, hashed and frozen. A mesh-convergence study on 40 of them fixes the mid mesh density (results within 0.15 °C of the finest). The split is 2,240 / 480 / 480, allocated by build-up family so that one family is absent from training entirely.
Trials:
Gaussian-process regression on 14 scalar geometry and material features: ±1.8 °C at the 95th percentile on validation, with almost all of the error above 45 mm cavity depth. This ruled out scalar feature encoding — it cannot represent the change in cavity convection regime.
Feed-forward network on the same features plus an explicit cavity Rayleigh-number term: ±0.9 °C, but a 0.4 °C spread across retraining seeds. The instability located the problem in the feature set rather than the architecture: the regime boundary remained under-determined.
Convolutional encoder over a rasterised section of the build-up, with the diurnal profile as a separate input channel: ±0.55 °C on validation, ±1.3 °C on the holdout, with most of the holdout error concentrated in the withheld build-up family.
Result reached: Inside the geometry envelope represented in training, roughly ±0.6 °C is achievable. Extrapolation to an unseen build-up family is not, and the year closes without establishing whether that limit is a data-volume problem or a formulation problem. The programme ends with a partial answer and an open question.
Where the boundary falls: Of roughly 2,100 engineering hours on the programme, about 380 sit in the hypothesis, the three trials and their evaluation; about 250 in generating and freezing the reference corpus; about 900 in parametric geometry, the batch pipeline, the interface and deployment to client teams; and about 570 in client project analyses, including compliance verification runs. On these illustrative facts, the 900 hours of tooling are treated as implementation using established methods where the relevant technical outcomes are assumed to be determinable in advance. The 570 hours of client analysis include activities associated with complying with statutory requirements or standards, which s 355-25(2) bars from being core. The 250 hours of reference-data generation exist only to feed the trials, so whether they attach as supporting activities turns on s 355-30, including dominant purpose where it is engaged. How each stream is treated in a registration and in the company's tax return is for the company and its registered tax agent to determine.
Examples of Useful Supporting Records
Simulation teams produce enormous volumes of output, and most of it records results rather than inquiry. Useful contemporaneous records may include (business.gov.au — record keeping for the R&DTI):
• The signed benchmark specification: envelope, boundary conditions, metric, target band and the pre-defined failure condition, dated before the trials;
• The search that established current knowledge did not already answer the question;
• Solver version, input-deck hashes and the mesh-convergence study fixing the reference configuration;
• The train / validation / holdout split manifest, with the rule used to allocate families;
• The experiment register: one row per model version, with configuration, metric outputs and the decision taken, including abandoned formulations;
• The decision log tying each observation to the next trial;
• The time-code mapping recorded as the work happened, separating experimentation, productisation and client delivery.
Where an RSP Fits
AusIndustry describes Research Service Providers as scientific or technical service providers you can engage to conduct R&D on your behalf, registered in specific research fields (business.gov.au). For an analysis team, an RSP earns its place before the sprint rather than after the year: fixing the envelope and the metric, freezing the corpus, designing the splits and the trial sequence so the record exists as the work happens.
Qualifying expenditure incurred to a non-associate RSP may still form part of the offset where total notional deductions are below the usual $20,000 threshold — see claiming R&D under $20,000. Using an RSP does not guarantee eligibility — you still self-assess.
Offset rates, the refundable and non-refundable tiers and the entitlement rules are covered separately in our refundable vs non-refundable offset article.
Frequently Asked Questions
Q: Is building a parametric or generative design tool eligible for the R&D Tax Incentive?
A: Building the tool is generally not the eligible activity: parametric geometry, batch pipelines and interfaces drive a validated solver whose method and technical result are already known, so the uncertainty is about effort. The distinct question is whether any component carried genuine uncertainty about achievable accuracy, convergence or optimisation performance that could only be resolved by systematic experiment. You self-assess.
Q: Is a machine-learning surrogate model for finite-element analysis R&D?
A: It may be, subject to self-assessment against all statutory requirements and exclusions, where the achievable error of the surrogate against validated reference solutions across a stated geometry envelope could not be known or determined in advance and was investigated through a systematic progression of work using an appropriate, documented evaluation method. A standard model trained on a well-understood parameter range, to an accuracy a competent professional could anticipate, generally will not be.
Q: Is automating a simulation workflow claimable?
A: Automating a validated solver workflow is generally not a core R&D activity, because the analysis method and its expected technical result are known and the work reduces labour. Automation built solely to execute or measure a genuine core experiment — for example a harness that runs a pre-registered trial sequence and computes the metric — may require separate assessment as a supporting activity, including the dominant-purpose test where it is engaged.
Q: Is analysis run to demonstrate compliance with a statutory requirement or standard eligible R&D?
A: Not as a core R&D activity. Section 355-25(2) of the ITAA 1997 excludes activities associated with complying with statutory requirements or standards from being core, and automating a verification or certification run does not change its character — unlike a surrogate whose accuracy against those same reference solutions was experimentally uncertain. Such a run may still qualify as a supporting R&D activity where it is directly related to a core activity and conducted for the dominant purpose of supporting it.
Sources & Further Reading
legislation.gov.au — Income Tax Assessment Act 1997 — Div 355, incl. ss 355-25 and 355-30
Related: property and construction · software and AI · what does not qualify · what an RSP is · claiming R&D under $20,000 · refundable vs non-refundable offset · R&D Tax Incentive in Adelaide
Talk to Ignition Research if you are planning analysis, simulation or automation work and need technical R&D support. As a Registered Research Service Provider at Lot Fourteen in Adelaide, we assist engineering consultancies and design-software teams with experimental design, technical R&D work and contemporaneous supporting records within our registered RSP scope. We do not determine R&DTI eligibility or provide tax advice: your company self-assesses and remains responsible for its own claim, with tax advice and lodgement handled by your tax adviser. Get in touch.
This article is general information from a Registered Research Service Provider about the R&D Tax Incentive. It is not tax, legal or financial advice; eligibility depends on your circumstances and you should self-assess and seek your own advice.
Thinking about a project like this?
If you're weighing up an AI, software or technical improvement project and can't tell yet whether it's implementation or research, start with a quick read on where it sits.