Tampere University AI in Civil Engineering Master's Thesis Guide

What this guide covers

Tampere University’s AI in Civil Engineering specialisation is now publicly documented as a current study option inside the Master’s Programme in Civil Engineering. It leads to the Master of Science (Technology) degree, is planned for two years, and has a total extent of 120 ECTS. The programme is organised by the Faculty of Built Environment at Hervanta Campus. The specialisation combines civil engineering, data science, statistics, computational modelling, machine learning and artificial intelligence, with explicit attention to reliability, uncertainty and engineering decision-making.

For thesis planning, the decisive current object is RAK.151 Master’s Thesis, Civil Engineering, 30 ECTS in the 2026–2027 Student’s Guide. It is an Advanced Studies course, available in English or Finnish, and graded on Tampere University’s general 0–5 scale. This guide therefore treats AI in Civil Engineering as a technology-field thesis under the Civil Engineering thesis route, not as a generic computing thesis.

The blocked programme page now exists

This guide was previously held back because Tampere’s applicant catalogue named AI in Civil Engineering but did not yet expose a stable programme-specific page. That evidence gap is now closed. The current page describes the degree, faculty, campus, 120 ECTS extent, learning focus, research environment and industry orientation. It also explains that students combine civil engineering expertise with data analytics, computational modelling, machine learning and AI rather than replacing engineering reasoning with generic data-science practice.

The programme page is therefore strong enough to establish the study-option identity and methodological environment. The Student’s Guide independently establishes the governing Civil Engineering thesis object. These are separate pieces of evidence and they agree.

The thesis is RAK.151, 30 ECTS

RAK.151 Master’s Thesis, Civil Engineering is the current thesis course for the Civil Engineering degree environment in 2026–2027. It carries 30 credits, belongs to Advanced Studies and uses the general 0–5 grading scale. The course is coordinated by Civil Engineering Studies in the Faculty of Built Environment and lists English and Finnish thesis completion options throughout the academic year.

Do not convert the thesis into 35 or 40 ECTS by adding other support courses. Introductory, information-search or research-method work may support the project, but the thesis itself remains 30 ECTS. Likewise, do not import a thesis code from Data Science, Electrical Engineering or another Tampere programme simply because the topic involves AI.

The degree is 120 ECTS, not a data-science conversion programme

The applicant-facing degree is Master of Science (Technology), 120 ECTS. Tampere describes the specialisation as educating civil engineers who can use computational and AI-based methods in engineering practice. The programme explicitly says its aim is not to turn civil engineers into data scientists, or data scientists into civil engineers. The point is to develop professionals who understand engineering systems and can judge when data-driven approaches genuinely improve decisions.

That distinction should shape the thesis. A machine-learning model is not automatically a civil-engineering contribution. The research question needs a built-environment, infrastructure, geotechnical, construction, real-estate or related engineering problem whose evidence can be improved through computation, sensing, modelling or AI.

Start from an engineering decision, not an algorithm

A strong thesis begins with an engineering question. Examples include predicting a geotechnical property, detecting construction anomalies, estimating deterioration, improving data flow across a construction process, modelling uncertainty in infrastructure decisions, classifying subsurface features, or evaluating whether an AI agent can support a bounded engineering workflow.

Avoid starting with “I will apply deep learning” before defining the engineering decision. The method should follow the problem. State what is being predicted, estimated, classified, optimised or supported; who or what uses the result; what baseline exists; and what error would be unacceptable in the real engineering context. This prevents technically impressive but practically irrelevant model development.

Reliability and uncertainty are central programme themes

The programme page explicitly highlights reliability and uncertainty. Civil engineering routinely deals with incomplete observations, variable materials, uncertain ground conditions, imperfect measurements and models that simplify reality. AI does not remove those uncertainties. It can introduce new ones through sampling, preprocessing, model selection, calibration and domain shift.

A thesis should therefore report more than a point estimate or accuracy score. Depending on the problem, use confidence intervals, prediction intervals, calibration analysis, sensitivity analysis, probabilistic modelling, repeated validation or scenario analysis. Explain whether uncertainty comes from measurement noise, natural variability, limited sample size, model assumptions or deployment conditions. An engineering recommendation should reflect those limits.

Data provenance must be traceable

Civil-engineering datasets may come from sensors, site investigations, BIM systems, construction logs, laboratory measurements, maintenance records, public geospatial sources, images, company databases or research projects. Document where the data came from, how it was collected, what units and coordinate systems apply, what period it covers, what quality-control steps were used and whether the collection process changed over time.

If data were supplied by a company, do not assume the dataset is complete or representative. Operational databases often reflect business processes rather than research sampling. Record missingness, duplicate records, coding changes, selection mechanisms and any preprocessing performed before you received the file.

Train, validation and test separation must reflect the engineering structure

Random row splitting can be invalid when observations are clustered by construction site, building, road segment, borehole, project, sensor, image sequence or time period. If data from the same physical asset appear in both training and test sets, performance can be exaggerated because the model has already seen closely related information.

Choose a split that represents the intended deployment. Group by project or asset when the model should generalise to new projects. Use temporal validation when predicting future conditions. Keep final test data untouched during feature selection and hyperparameter tuning. Any preprocessing that learns from the data should be fitted inside the training workflow, not on the full dataset.

Compare AI against defensible engineering baselines

A thesis should not claim improvement merely because a neural network outperforms a deliberately weak baseline. Compare the proposed method with established engineering calculations, simple statistical models, current practice or other reasonable computational approaches when those are available.

Define the comparison before final evaluation. Use the same data partitions and performance criteria for all models. If the complex model improves average error only slightly while becoming harder to interpret, more expensive to run or less stable under distribution shift, discuss whether the improvement is engineeringly meaningful.

Metrics must reflect consequences

Accuracy, RMSE, R-squared or F1 are not interchangeable. Choose metrics that correspond to the engineering decision and error cost. A model used to flag unsafe conditions may need sensitivity and false-negative analysis. A cost-estimation model may need absolute and relative error. A probabilistic prediction may need calibration as well as sharpness.

Report class imbalance and threshold choices. Where a decision threshold is used, explain how it was chosen. Do not optimise the threshold on the same final test set used to report performance. If different error types have different safety or financial consequences, make that asymmetry visible.

AI2Geo shows why measurement validity matters

The current programme page highlights AI2Geo, a project combining advanced imaging and AI to improve determination of subsurface ground properties. The research portal describes a multidisciplinary experimental programme linking micro-level mineral, chemical, microbial and imaging information with physical soil characteristics.

This is a useful methodological lesson. An AI model is only as meaningful as the measurement chain that produces its inputs and labels. When using imaging or sensor-derived features, document calibration, spatial resolution, preprocessing and the reference measurement used as ground truth. A model can reproduce systematic measurement bias very accurately.

AI Champion shows why workflow claims need bounded evaluation

The programme also points to AI Champion, which develops AI agents intended to improve the flow of information from organisational data silos into engineering practice. A thesis in this area should distinguish a functioning prototype from a validated workflow improvement.

Measure the outcome that matters. If the claim is faster information retrieval, record time and task success. If the claim is better decision support, define decision quality and compare against current practice. If the system uses an LLM, evaluate hallucination, omission, retrieval failure, prompt sensitivity and source traceability. Demonstrating that an agent can produce fluent text is not evidence that it improves engineering decisions.

Generative AI requires source-grounded verification

LLM-based tools can help with information extraction, classification, coding assistance or document workflows, but they can also fabricate facts, citations and calculations. In a civil-engineering thesis, generated outputs must not be treated as authoritative merely because they sound plausible.

Define what sources the system is allowed to use and how outputs are checked. For retrieval-augmented systems, measure retrieval as well as answer quality. Keep protected company information out of external services unless an approved basis exists. Tampere’s thesis guidance requires AI use to be agreed with the primary supervisor and handled according to current university instructions.

Simulation is not field validation

Synthetic data, simulations and digital twins can be valuable, especially when real failures or rare events are hard to observe. But performance on simulated data does not prove field performance. Explain which assumptions generate the simulated environment and which real-world effects are omitted.

If a model is trained partly or entirely on simulated data, test it on independent real measurements where feasible. When that is not possible, narrow the conclusion to simulation performance and identify the validation still required before engineering use.

Sensor and monitoring studies need calibration and time alignment

For structural, geotechnical or environmental monitoring, document the sensor model, units, sampling rate, installation conditions, calibration, drift handling and synchronisation. Missing or misaligned timestamps can create false relationships. If sensors are replaced during the study, record the change.

Filtering and smoothing are analysis choices, not invisible cleaning. Report them. If a derived feature depends on window length or threshold settings, test whether conclusions are robust to reasonable alternatives.

Spatial data require spatial validation

Geotechnical, infrastructure and urban built-environment data are often spatially autocorrelated. Nearby samples may be more similar than distant samples. A random split can therefore overstate generalisation.

Consider spatial blocking, leave-location-out validation or another approach that reflects the intended use. State coordinate reference systems and transformations. If data are aggregated to grid cells, segments or zones, explain the aggregation and avoid inferring individual-level behaviour from area-level patterns.

Interpretability must match the engineering use

Some civil-engineering decisions require not only prediction but also a reason that can be inspected. Feature importance, partial-dependence plots, SHAP-style explanations or simpler surrogate models can help, but none should be treated as automatic causal explanations.

If interpretability is part of the thesis claim, validate the explanation method. Check whether explanations are stable across resampling and whether domain experts find them consistent with the physical problem. A plausible-looking explanation from a black-box model can still be unstable or misleading.

Domain shift is an engineering risk

Models can fail when soil type, sensor hardware, building stock, climate, contractor practice or data-collection protocol changes. Performance on one project is therefore not proof of portability to another project or country.

Describe the training domain and intended deployment domain. Where possible, test on a genuinely separate site or project. If external validation is unavailable, state that limitation explicitly and avoid using language such as “generalisable” or “ready for deployment” without evidence.

Company thesis projects need an academic contribution

The programme promotes strong company involvement and real-world thesis opportunities. A commissioned thesis can be excellent, but the company’s practical objective and the academic research question are not identical.

Agree early on what can be published, which data can be retained, who owns code or models and how negative findings will be handled. The final thesis must still explain methods and evidence clearly enough for academic assessment. Confidential operational details should be separated from the public manuscript where necessary rather than making the thesis scientifically opaque.

Research ethics depends on the actual method

Many AI in Civil Engineering projects use technical data and may not involve human participants. Others can include interviews, worker observations, location traces, user studies or personal data. Tampere’s Ethics Committee of the Tampere Region handles non-medical human-sciences review under the applicable criteria, including technological research involving human participants.

Do not state that every AI thesis needs ethical review. The correct rule is to classify the actual design with the supervisor before data collection. If personal data are processed, Tampere’s student data-protection guidance requires the processing purpose, scope, storage, retention and lawful basis to be planned in advance.

Personal and commercially sensitive data need explicit controls

Construction and property datasets can contain names, contact details, access logs, location histories, photographs, work schedules or other information linked to individuals. They can also contain commercially confidential cost, contract or operational information.

Apply data minimisation. Separate identifiers when they are not analytically needed. Use approved storage and access controls. If the thesis is part of a larger university project, clarify whether the University or another organisation is the data controller. Do not upload protected data into external AI services merely for convenience.

RAK.300 is research preparation, not the thesis

The current RAK.300 Individual Research Work in Civil Engineering course is 2–5 ECTS and develops planning, literature work, research or design methods, analysis and report writing. It can be useful preparation, but it is not RAK.151 and should not be presented as thesis credit.

Similarly, RAK.250 Introduction to Civil Engineering Studies supports programme orientation and information-search skills. Keep these course objects separate when describing the 120 ECTS structure.

The technology thesis process has a supervision plan

Tampere’s technology/architecture thesis instructions require a Thesis Supervision Plan prepared with the supervisor. Use it to define the research question, milestones, meeting practices, data access, validation plan, expected outputs and timetable.

For computational work, also freeze evaluation logic early: baseline, data split, primary metrics and acceptable changes. This reduces the risk of repeatedly changing the method until a preferred result appears.

Maturity test, Turnitin and Trepo

The technology thesis process includes a maturity test. For international Master’s students, Tampere states that the thesis abstract serves as the maturity test, with the examiner assessing its content and no language checking. Other students should follow the current maturity-test rules that apply to their prior language demonstration.

Before formal examination, the thesis undergoes a Turnitin originality check. After supervisor permission, the final thesis is deposited through Trepo, Tampere University’s institutional repository. Turnitin checks textual similarity; it does not validate code, statistics, engineering assumptions or AI performance.

Public thesis and protected material

Trepo publication does not mean every raw dataset, source file or company appendix should be public. Plan the public version from the beginning. Use aggregated or anonymised results where appropriate and keep protected source data in approved environments.

A public thesis should still explain enough about the data, model, validation and limitations for the examiner and reader to understand the research. “Confidential” should not become a substitute for methodological transparency.

Final AI in Civil Engineering checklist

Confirm that your Sisu plan still shows RAK.151, 30 ECTS, 0–5 for the thesis. Define the engineering decision before choosing the AI method. Document data provenance, measurement quality, grouping and temporal structure. Prevent train-test leakage. Compare against defensible baselines. Report uncertainty, calibration and failure cases. Keep simulation claims separate from field evidence. Map human-participant, personal-data and company confidentiality requirements before collection. Agree AI use with the supervisor, maintain a supervision plan, complete Turnitin, the maturity-test route and Trepo submission according to current Tampere instructions.

The strongest thesis is not the one with the most complex model. It is the one that makes a clear engineering claim, uses a defensible method and shows exactly how far the evidence supports the conclusion.

Reproducibility should connect code, data and engineering assumptions

A computational thesis should make it possible to reconstruct how raw records became the reported engineering result. Record dataset version, filtering, feature construction, model configuration, software and package versions, random seeds, training environment and the exact evaluation split. Use version control for analysis code and preserve a run record that links a result table or figure to the configuration that produced it.

Reproducibility does not require publishing protected company data. When raw data cannot be released, provide a data dictionary, transformation logic, model configuration and, where feasible, synthetic or non-sensitive examples. The aim is to make the analytical reasoning inspectable without violating confidentiality.

Missing data and exclusions are part of the result

Civil-engineering datasets often contain missing sensors, failed measurements, incomplete site records or observations excluded by quality criteria. Report how much data were available at each stage and why observations were removed. If missingness is related to difficult site conditions or system failure, complete-case analysis can bias the result toward easier cases.

Predefine exclusion rules where practical. If rules change during troubleshooting, record the change and test whether the main conclusion depends on it. Do not silently remove difficult observations because they reduce model performance.

Physical plausibility still matters for data-driven models

A high-scoring model can produce outputs that violate engineering constraints. Where relevant, check units, sign conventions, monotonic relationships, conservation principles, feasible ranges and known boundary conditions. Compare predictions with expert expectations and investigate large residuals rather than treating them only as statistical noise.

If a model contradicts established physical understanding, that can be an important finding, but it needs investigation. The thesis should distinguish genuine new evidence from a data error, leakage, confounding or model artefact.

Manuscript freeze should include an evidence audit

Before submission, trace every major conclusion back to its data, model run and validation result. Label whether the conclusion is descriptive, predictive, simulation-based, field-validated or deployment-oriented. Remove wording that goes beyond the tested domain. Re-run at least one key analysis from a clean environment where feasible and check that tables, figures and manuscript numbers agree.

This final audit is particularly important for AI projects because small changes in preprocessing, random seeds or data partitions can alter results. The final thesis should report the evidence that actually survived verification, not the most favourable result produced during experimentation.

Treat negative and null results as engineering evidence

An AI model that fails to outperform a simple baseline can still produce a strong thesis if the study explains why. A negative result may reveal that available measurements do not contain enough signal, that labels are unreliable, that domain shift is too large or that the existing engineering method is already difficult to improve. Report these outcomes instead of searching repeatedly for a favourable configuration. Distinguish planned analysis from exploratory troubleshooting, preserve failed runs when they inform the argument, and explain what additional evidence would be needed before the proposed method should be reconsidered for practice.