PhD research topics in engineering

A PhD topic is a question, not a title. Each card states the problem, what a successful thesis would produce, why nobody has settled it yet and a first experiment small enough to try in a few weeks.

Which PhD research topics in engineering are open in 2026?

The 20 below are active areas where published work still disagrees or leaves a clear gap. Read the card, then search the last two or three years of papers on it before you take it to a supervisor.

AI and data science

Sixteen more computing topics, each with recent surveys and a public dataset, are in our PhD topics in computer science.

• Topic idea · CSE · AIConfidence that stays honest when the data changesProblem: A model’s stated confidence stops matching its accuracy once real data drifts away from the training data.Expected output: A calibration method, or a test that warns when calibration has failed, validated across several kinds of measured shift.
  Why it is open, and a first step

  Why it is open: most calibration methods are tuned and checked on data that resembles the training set
  First step: reproduce two calibration methods on a benchmark with known shifts and chart where each breaks down.

• Topic idea · CSE · AIAdapting large vision models to inspection with few labelsProblem: Large pre-trained vision models are trained on everyday photographs, while factories have few labelled images of their own parts.Expected output: An adaptation method that reaches useful accuracy from a handful of labelled images, with its limits mapped across product types.
  Why it is open, and a first step

  Why it is open: it is not yet clear which adaptation methods transfer to textures and defects that look nothing like web images
  First step: compare three adaptation methods on one public industrial dataset as the number of labels falls.

• Topic idea · CSE · AICausal effect estimates that decision makers can trustProblem: Predictive models show what is associated with an outcome, not what would change it, and decisions need the second.Expected output: A method for estimating the effect of an action from observational records, with a way to test how sensitive it is to hidden causes.
  Why it is open, and a first step

  Why it is open: the assumptions behind causal estimates cannot be checked from the data alone
  First step: apply two estimators to a dataset where a randomised result is also available, and compare.

• Topic idea · CSE · AISynthetic data with a privacy level that can be auditedProblem: Organisations want to share tabular data for research, but removing names does not stop people being re-identified.Expected output: A generator for tabular data with a stated privacy level, an audit that tests it, and the loss in usefulness measured.
  Why it is open, and a first step

  Why it is open: stronger privacy lowers the value of the data, and the trade-off depends heavily on the dataset
  First step: train one private generator on a public census-style table and measure both attack success and model accuracy.

• Topic idea · CSE · AIMaking a trained model forget one person’s dataProblem: Privacy law gives people the right to have their data removed, but retraining a large model for every request is too costly.Expected output: An unlearning method with a test that shows the data’s influence has gone, at a fraction of the cost of retraining.
  Why it is open, and a first step

  Why it is open: there is no agreed way to prove that a model has forgotten
  First step: implement one approximate method and compare it with full retraining using a membership test.

Networks, security and systems

• Topic idea · CSE · ITLearning-based control for open radio access networksProblem: Open radio access networks let software from different vendors control the radio, and hand-written rules cannot keep up with changing traffic.Expected output: A learned controller for allocating radio resources that respects safety limits while it learns.
  Why it is open, and a first step

  Why it is open: a controller that explores freely can disrupt a live network, so learning has to stay inside constraints
  First step: build a simulated cell with an open-source stack and compare a rule-based and a learned scheduler.

• Topic idea · CSE · ITIntrusion detectors that resist deliberate evasionProblem: Machine learning detectors can be evaded by attackers who alter their traffic slightly.Expected output: An attack model limited to changes that keep traffic valid, and a detector hardened against it.
  Why it is open, and a first step

  Why it is open: many published attacks change features in ways that real packets could not carry
  First step: take a public intrusion dataset, define which features an attacker can really change, and measure evasion.

• Topic idea · CSE · ITCarbon-aware scheduling of machine learning jobsProblem: Training runs draw the same power whether the grid is running on coal or on solar.Expected output: A scheduler that moves flexible jobs in time or place to cut emissions, with the delay it causes measured.
  Why it is open, and a first step

  Why it is open: the saving depends on carbon forecasts that are themselves uncertain
  First step: replay a public cluster trace against regional grid carbon data and measure what simple shifting saves.

• Topic idea · CSE · ITKeeping a digital twin in step with limited telemetryProblem: A digital twin of a network or a plant drifts from the real system when measurements are sparse or late.Expected output: A method that decides which measurements to collect to keep the twin accurate enough for control.
  Why it is open, and a first step

  Why it is open: collecting everything is too costly, and it is unclear how little data is enough
  First step: build a twin of a small simulated system and measure its error as telemetry is thinned out.

Electronics and communication

• Topic idea · ECE · EnTCOne waveform for both sensing and communicationProblem: Future radios are expected to detect objects and carry data with the same signal, and the two goals pull the design in different directions.Expected output: A waveform or beamforming design with the trade-off between sensing accuracy and data rate characterised.
  Why it is open, and a first step

  Why it is open: improving one of the two usually costs the other, and the best balance depends on the scene
  First step: simulate a single link with one target and plot detection performance against data rate.

• Topic idea · ECE · EnTCChannel estimation for reconfigurable intelligent surfacesProblem: A reflecting surface with hundreds of passive elements cannot measure the channel itself.Expected output: An estimation method whose pilot overhead grows slowly with the number of elements, with accuracy bounds.
  Why it is open, and a first step

  Why it is open: pilot overhead rises with the size of the surface, and passive elements cannot process signals
  First step: reproduce a baseline estimator in simulation and measure overhead as the surface grows.

• Topic idea · ECE · EnTCApproximate arithmetic for low-energy AI chipsProblem: Exact arithmetic spends energy on precision that a neural network does not always need.Expected output: A way to choose where approximate circuits are safe in an accelerator, with accuracy loss and energy saved reported together.
  Why it is open, and a first step

  Why it is open: the tolerable error differs from layer to layer and from task to task
  First step: replace the multipliers of one network layer with approximate ones in simulation and measure both effects.

• Topic idea · ECE · EnTCTraining on a microcontroller, not only inferenceProblem: Devices in the field meet conditions that their factory-trained model never saw, and they have kilobytes of memory, not gigabytes.Expected output: A training method that adapts a model on the device within its memory and energy budget.
  Why it is open, and a first step

  Why it is open: standard training stores far more intermediate data than a microcontroller can hold
  First step: adapt only the last layer of a small model on a development board and measure memory and accuracy.

• Topic idea · ECE · EnTCCuffless blood pressure that works across peopleProblem: Estimates of blood pressure from optical pulse sensors drift over time and differ from person to person.Expected output: An estimation method tested on people left out of training, with its error reported against a clinical standard.
  Why it is open, and a first step

  Why it is open: models that fit one person well often fail on another, and many studies test on the same people they train on
  First step: evaluate a published model on a public dataset with a strict split by person.

Electrical engineering and energy

• Topic idea · EEE · EnergyBattery health estimates that transfer to real useProblem: Models of battery ageing are built from controlled laboratory cycling, while real vehicles are driven irregularly.Expected output: A state-of-health estimator that holds up across usage patterns and cell types, with its uncertainty stated.
  Why it is open, and a first step

  Why it is open: laboratory data and field data differ, and labelled field data is scarce
  First step: train on one public cycling dataset and test on another with a different cell or protocol.

• Topic idea · EEE · EnergyStable control of microgrids with a high solar shareProblem: Grids fed mainly through inverters respond faster and less predictably than grids with heavy rotating machines.Expected output: A data-driven controller with a stability argument, tested in simulation under sudden changes in sun and load.
  Why it is open, and a first step

  Why it is open: learned controllers perform well on average but are hard to certify as stable
  First step: model a small microgrid and compare a conventional controller with a learned one under step changes.

• Topic idea · EEE · EnergyAppliance-level energy use from ordinary meter readingsProblem: Working out which appliance used how much energy needs fast sampling that utility meters do not provide.Expected output: A disaggregation method for slow meter readings, with its accuracy for each appliance and its limits made clear.
  Why it is open, and a first step

  Why it is open: at slow sampling rates the signatures of different appliances overlap
  First step: downsample a public household dataset and measure how accuracy falls for each appliance.

Applied AI in other fields

• Topic idea · Mechanical · Civil · AgriculturePhysics-informed models for structural health monitoringProblem: Bridges and machines rarely fail, so there is little failure data to learn from.Expected output: A model that combines the governing physics with sensor data to detect damage earlier than a model using data alone.
  Why it is open, and a first step

  Why it is open: it is unsettled how to weigh physics against data when the physical model is itself imperfect
  First step: fit a physics-informed model to a simulated beam with known damage, then to a public vibration dataset.

• Topic idea · Mechanical · Civil · AgricultureCrop yield prediction that transfers between districtsProblem: A yield model trained where records are good performs poorly where they are sparse.Expected output: A method that carries a model from data-rich to data-poor districts using satellite and weather data, with the error mapped.
  Why it is open, and a first step

  Why it is open: ground truth is available only as coarse official figures, and crops and practices differ by region
  First step: build a baseline from public satellite indices and official yield figures for one state.

• Topic idea · Mechanical · Civil · AgricultureTraffic signal control for mixed trafficProblem: Signal control methods assume vehicles keep to lanes, which does not hold on roads shared by two-wheelers, cars, buses and pedestrians.Expected output: A control policy trained and tested on mixed traffic, compared with fixed-time and actuated signals.
  Why it is open, and a first step

  Why it is open: most simulators and published results assume lane discipline
  First step: calibrate a traffic simulator to video counts from one junction and compare three control policies.

How do you turn a problem into a research gap?

A problem says what is wrong. A gap says what is missing from the published attempts to fix it. Get from one to the other in four moves.

1. Read the recent surveys and list what each calls an open issue.
2. Tabulate the main methods: the data each used, what it measured and what it left out.
3. Find the empty column. A setting nobody tested, an assumption nobody checked or a comparison nobody ran.
4. State the gap in one sentence that names what exists and what does not.

That sentence becomes the centre of the proposal. Our guide to the research proposal for PhD admission shows where it goes and what surrounds it.

How do you test a topic before you commit?

What comes after choosing a topic?

• The proposal, for admission or the first review: the problem, the gap, your questions, a method and a timeline.
• The coursework and the literature review, which sharpen the gap.
• The synopsis, for registration: see the PhD synopsis format.
• Experiments, papers and the thesis, in the order your university’s regulations set.

For support at each stage, see PhD guidance and assistance. Doing an M.Tech first? See the M.Tech project ideas, which are smaller versions of the same kind of question.

Frequently asked questions