Convergences
New research often comes from joining two things nobody has joined before. Here are 20 such ideas, each taking a method that already works in one field and applying it to a concrete problem in another, and each checked to make sure nobody has done it yet. Every idea says what the problem is, what gets borrowed, and what it would give.
- 103pairings of two fields' methods and problems
- 48left after a first count of papers, searched by hand
- 20untouched and most promising, shown below
- #1 8.5
Does phishing training work? Measure it the way flu-vaccine studies do
- The problem
- Companies spend heavily on anti-phishing training but cannot tell whether it helps. Comparing trained with untrained staff misleads, because the people who report suspicious emails are already the careful ones.
- What to borrow
- Flu-vaccine studies face the same trap: people who get tested differ from people who do not. Their fix, the test-negative design, compares only people who came in for a test: among those who tested positive and those who tested negative, how many were vaccinated? Do the same with the emails staff report: among reported emails that were real attacks and those that were harmless, how many reporters had been trained?
- What it gives
- A cheap, routine estimate of how much training cuts successful attacks, from data security teams already collect.
Status. Not done. Studies of phishing training use randomised trials or click rates on fake phishing emails; none uses this design. Closest work: Are You at Risk? Profiling Organizations and Individuals Subject to Targeted Attacks (Thonnard et al.) · Understanding the Efficacy of Phishing Training in Practice (2025)
Untouched 9Fields apart 8Real method 9Payoff 8
- #2 8.0
Find power-grid faults with the error-finding algorithms of quantum computers
- The problem
- When a line in a local power grid fails, only a few sensors and protection switches raise alarms. Operators must work out from this sparse, sometimes misleading pattern where the fault is, and in large networks that is slow.
- What to borrow
- Quantum computers face the same puzzle. Their errors are invisible and only some check points light up, so engineers built decoders that find the most likely errors from those lights, on graphs with millions of nodes, in a fraction of a second.
- What it gives
- Faster, more reliable fault location from the alarms grids already have, with known conditions under which the answer is the best possible.
Status. Not done. Grid fault location uses state estimation and machine learning; nobody has tried quantum decoders. Closest work: Sparse Blossom: correcting a million errors per core second with minimum-weight matching (2023)
Untouched 9Fields apart 8Real method 8Payoff 7
- #3 8.0
Sound pulses that deepen sleep, with a guarantee against waking the sleeper
- The problem
- Soft clicks timed to the brain's slow waves during deep sleep can strengthen it and improve memory; the technique is being tested in older adults and patients. The main risk is waking the person, and devices handle it with fixed volume rules.
- What to borrow
- Robots and self-driving cars use control barrier functions: a mathematical filter that checks every action and blocks any that would push the system into a forbidden zone, such as hitting an obstacle. Here the forbidden zone is waking up, read from the live brain signal.
- What it gives
- Stimulation as strong as each person tolerates, which may give more benefit, backed by a guarantee instead of a rule of thumb.
Status. Not done. Current systems use fixed or simple adaptive volume rules. Closest work: Driving Sleep Slow Oscillations by Auditory Closed-Loop Stimulation - A Self-Limiting Process (2015) · Closed-loop auditory stimulation of sleep slow oscillations: Basic principles and best practices (2023)
Untouched 9Fields apart 7Real method 8Payoff 8
- #4 7.8
Reconstruct hidden chains of infection with quantum error-correction decoders
- The problem
- In an outbreak, health officials see only the people who test positive. Who infected whom, through which unseen people, stays hidden, and the tools that infer it are accurate but slow.
- What to borrow
- Quantum decoders solve a look-alike problem: they see only the two ends of each chain of errors and must infer the chain in between, fast. Positive tests play the visible ends; the contact network plays the chip's layout.
- What it gives
- Near-instant estimates of likely transmission routes on large contact networks, to decide whom to trace first.
Status. Not done. Transmission-tree inference uses slow Bayesian sampling; nobody has tried quantum decoders. Closest work: Experimentally Informed Decoding of Stabilizer Codes Based on Syndrome Correlations (2025) · Novel Decoding Algorithm for Noiseless Non-Adaptive Group Testing (2026)
Untouched 9Fields apart 9Real method 7Payoff 6
- #5 7.8
Personalise the bedtime schedule in insomnia therapy with the controller that runs chemical plants
- The problem
- The most effective insomnia treatment first limits time in bed, then extends it by a fixed 15 minutes a week if sleep improves. Everyone gets the same schedule, so some patients progress too slowly and others give up.
- What to borrow
- Model predictive control, used in chemical plants and building climate systems, looks a few steps ahead with a model of the system and picks the best next move within set limits. The model here is the standard one of sleep pressure and body clock, fitted to each patient's sleep diary; a limit is, for example, at least five hours in bed.
- What it gives
- A schedule tuned to each patient's response, which could shorten treatment and cut dropouts.
Status. Not done. A patent and a few studies personalise the schedule with rules, not with a predictive controller. Closest work: SYSTEM AND METHOD FOR PERSONALIZATION OF SLEEP RESTRICTION THERAPY (patent application US20220110582) · Optimization of light exposure and sleep schedule for circadian rhythm entrainment (2021)
Untouched 8Fields apart 7Real method 8Payoff 8
- #6 7.8
Estimate how many AI test questions leaked into training data, the way antibody surveys do
- The problem
- An AI benchmark means something only if the model never saw the test questions in training. The tools that detect such leaks err in both directions, so reported leak rates are biased and come without honest error bars.
- What to borrow
- Antibody surveys have the same issue: blood tests miss some past infections and flag some people who were never infected. Epidemiologists correct the raw positive rate with the test's known error rates, using a standard formula, and report a confidence interval.
- What it gives
- Corrected, comparable leak estimates for each benchmark, so readers know how far a score can be trusted.
Status. One 2026 paper mentions the formula in passing; nobody has measured the detectors' error rates and applied the correction. Closest work: When Is Benchmark Contamination Detectable? Information Limits and Power-Calibrated Audits (2026) · Correcting test set contamination by spiking the training data (2026)
Untouched 6Fields apart 8Real method 9Payoff 8
- #7 7.5
Ask night owls and early birds differently when the grid needs people to use less power
- The problem
- Utilities pay households to use less power at peak hours. Many households opt out because the requested hours clash with their routine, especially their sleep.
- What to borrow
- Sleep science measures chronotype, whether someone is a morning or an evening person, with a short questionnaire, and it may also show in meter data. Schedule each household's requests around its chronotype.
- What it gives
- Fewer opt-outs and less discomfort for the same relief to the grid.
Status. Not done. Studies look at comfort and daily routines in general, not at chronotype. Closest work: Assessing thermal comfort and participation in residential demand flexibility programs (PNNL, 2026) · Nambiar 2024 residential DR case studies (CBE Berkeley)
Untouched 9Fields apart 8Real method 6Payoff 7
- #8 7.5
Spot flu waves early from household electricity use
- The problem
- Flu surveillance relies on doctor visits and lab tests, which lag the real spread by one to two weeks.
- What to borrow
- Sick people stay home on weekdays, so daytime electricity use rises. Smart meters record this every 15 minutes for whole towns.
- What it gives
- A free, near-real-time signal that could warn of flu waves days before clinics do.
Status. Floated in a 2016 trade column and linked to COVID lockdowns, but never checked against official flu data. Closest work: Tracking the Bugs with Smart Meter Data (Public Utilities Fortnightly column, 2016) · Electricity consumption reveals proactive community response to COVID-19 progression (NUS, 2021)
Untouched 7Fields apart 8Real method 7Payoff 8
- #9 7.3
Date the rise and burst of research hype with the test economists use for stock bubbles
- The problem
- Research topics go in and out of fashion, but hype is called by gut feeling, usually after the fact.
- What to borrow
- Economists have a statistical test (GSADF) that tells when a price grows faster than any normal process would, and dates when the bubble began and when it burst. Apply it to the yearly number of papers on a topic.
- What it gives
- An objective, early flag for overheated topics, useful to funders, editors and young researchers choosing a direction.
Status. Not done. The test has been applied to AI stocks, not to publication counts. Closest work: Examining the Relationship between Scientific Publishing Activity and Hype-Driven Financial Bubbles: A Comparison of the Dot-Com and AI Eras (2025) · Limited Diffusion of Scientific Knowledge Forecasts Collapse (2023)
Untouched 9Fields apart 6Real method 8Payoff 6
- #10 7.3
Are babies near-perfect learners? Compare them with the theoretical minimum
- The problem
- Psychologists measure how many repetitions infants need to pick words out of a stream of speech, but have no yardstick for whether that number is impressive.
- What to borrow
- Learning theory in computer science proves the minimum number of examples any learner, human or machine, needs for a given task. Work out that minimum for the same word-finding tasks used in infant experiments.
- What it gives
- If babies come close to the minimum, they learn near-optimally; if they need fewer examples than it allows, they must bring built-in knowledge. That puts numbers on a long nature-versus-nurture argument.
Status. Not done. Closest work: Grammatical inference and first language acquisition (Clark, 2004) · Testing the limits of statistical learning for word segmentation (Johnson & Tyler, 2010)
Untouched 8Fields apart 8Real method 7Payoff 6
- #11 7.3
Manage malware detectors the way hospitals manage antibiotics
- The problem
- When one malware detector is used everywhere, attackers learn to slip past it and it loses power over time. Security teams have no model of how fast this happens or how to slow it.
- What to borrow
- Overusing an antibiotic breeds resistant bacteria in the same way. Medicine has numerical models of how fast resistance grows with use, and of stewardship policies, such as rotating or restricting drugs, that slow it.
- What it gives
- A way to plan how detectors are deployed so they stay effective for longer.
Status. The analogy is made informally; nobody has fitted resistance models to malware data. Closest work: MalTree: Tracing Malware Evolution from Embeddings at Scale (2026) · Measuring the Speed of the Red Queen's Race; Adaption and Evasion in Malware (Sophos talk)
Untouched 8Fields apart 8Real method 7Payoff 6
- #12 7.3
Do AI language models see intention where there is none, as people do?
- The problem
- A leading theory of why people believe in unseen spirits says our minds over-detect agents: we hear a rustle and assume someone is there. It is hard to test, because every person's upbringing and culture are mixed in.
- What to borrow
- Give large language models the standard story-based tests psychologists use for this bias. The models learned only from human text, with no body and no survival instinct.
- What it gives
- If models show the same bias, the bias may travel in language itself rather than in evolved instinct, a new test of a major theory. AI builders also learn whether their models read intent into accidents.
Status. Only an unpublished thesis tests a related bias; this test has not been run. Closest work: Praca licencjacka [Univ. Warsaw bachelor thesis on teleological bias in Gemini models] · The Double-Edged Sword of Anthropomorphism in LLMs (Reinecke et al., 2025)
Untouched 7Fields apart 9Real method 7Payoff 6
- #13 7.3
Watch AI models in use with the alarm systems built for disease outbreaks
- The problem
- AI models in use slowly get worse as the world changes. Companies watch error counts with simple thresholds, which either raise false alarms over normal weekly and seasonal ups and downs or react late.
- What to borrow
- Public-health agencies have algorithms, tested for decades, that flag unusual weekly case counts while allowing for seasons and trends. A model's error counts are the same kind of data.
- What it gives
- Fewer false alarms and earlier real ones, using methods agencies already trust.
Status. Monitoring uses simpler tools; nobody has applied these outbreak detectors. Closest work: Detecting and Monitoring Bias for Subgroups in Breast Cancer Detection AI (2025) · Monitoring Machine Learning Models: Online Detection of Relevant Deviations (2023)
Untouched 6Fields apart 8Real method 8Payoff 7
- #14 7.0
Predict which parts of a neural network will specialise or go dead during training, with a tool from cell biology
- The problem
- During training, some units of a network become specialised and others go dead. Usually we see this only afterwards.
- What to borrow
- Biologists can tell what a cell is turning into from a single snapshot, by comparing its newly made and its mature RNA (RNA velocity). Two nearby training checkpoints give the same kind of before-and-after reading for each unit.
- What it gives
- Early warning of dead units and a map of how features form, useful for cheaper training and for understanding models.
Status. Not done. Closest work: The Early Phase of Neural Network Training (2020)
Untouched 9Fields apart 7Real method 6Payoff 6
- #15 7.0
Reuse AI models trained on cancer slides to analyse metal under the microscope
- The problem
- Engineers judge metal quality from microscope images of grains, boundaries and defects, but labelled images are scarce, so AI tools for them are weak.
- What to borrow
- Very large AI models have been trained on millions of stained tissue slides, and tissue and etched metal look surprisingly alike under a microscope: grains, boundaries, inclusions.
- What it gives
- A strong starting model for materials images, and almost free to try.
Status. Not done. Materials AI starts from everyday photos or its own small datasets. Closest work: Machine Learning Models Pre-trained on MicroNet Show Improved Segmentation and Microstructure Quantification (IMAT 2023) · Addressing materials' microstructure diversity using transfer learning (2021)
Untouched 9Fields apart 6Real method 6Payoff 7
- #16 7.0
Diagnose sleep apnea from a sample of the night, with a guaranteed error rate
- The problem
- Diagnosing sleep apnea needs a technician to score a whole night of recordings. It takes hours and causes long waiting lists.
- What to borrow
- A branch of computer science, property testing, decides whether data has a property by reading only a small random sample, with a proven bound on the chance of being wrong.
- What it gives
- Faster screening with a stated error guarantee, rather than a hope that the sample is typical.
Status. Scoring only part of the night has been tried, but without any guarantee. Closest work: Polysomnography scoring for sleep apnea: use of a sampling method · Prerau lab AHI uncertainty tools
Untouched 7Fields apart 8Real method 7Payoff 6
- #17 6.8
Find the spots in a social network where information gets trapped, with a trick from physics
- The problem
- Some messages never spread beyond a small group, even in a well-connected network. This is usually blamed on the content or on people's interests.
- What to borrow
- In certain crystal lattices physicists found waves trapped on a handful of sites, because they cancel out everywhere else. The same mathematics may apply to the shape of a social network.
- What it gives
- An explanation for stuck information that comes from the network's structure, not its content, and a way to find those spots in advance.
Status. Not done. Closest work: Localization Phenomena in Large-Scale Networked Systems: Robustness and Fragility of Dynamics (2024) · Compact localized states and flatband generators in one dimension (2016)
Untouched 9Fields apart 6Real method 7Payoff 5
- #18 6.8
When does an AI prediction with a guarantee count as justified belief?
- The problem
- A popular AI method, conformal prediction, promises that its answers are right, say, 90% of the time. Users read this as trustworthiness for each case, but the promise holds only on average.
- What to borrow
- Philosophers hold that a belief is justified if it comes from a reliable process, and they know the theory's weak spot: what counts as reliable depends on how cases are grouped. That weak spot is exactly the gap between an on-average and a per-case guarantee.
- What it gives
- A clear account of what such guarantees do and do not license, for philosophers and for people using them in medicine or law.
Status. Not done. Closest work: Beyond Accuracy: Epistemic Justification in Trustworthy Machine Learning (2026, ICML) · When is a Prediction Knowledge? (2019)
Untouched 8Fields apart 8Real method 6Payoff 5
- #19 6.8
Read your body clock from a smartwatch with an AI that obeys the sleep equations
- The problem
- Sleep science has a classic model: sleep pressure builds while you are awake, and a body clock opens and closes the gate to sleep. Fitting it to one person needs clean sleep diaries; smartwatch data is noisy and full of gaps.
- What to borrow
- Physics-informed neural networks learn from data while being forced to obey known equations, which lets them cope with noisy, patchy measurements.
- What it gives
- Personal estimates of sleep pressure and body-clock timing from everyday wearables, for shift planning or jet-lag advice.
Status. Not done. Current fits use classical statistics. Closest work: Sleep timing estimation from smartphone context information (Wahl & Amft) · A method for characterizing daily physiology from widely used wearables (2021)
Untouched 8Fields apart 7Real method 6Payoff 6
- #20 6.8
Pack chatbot memory on GPUs with the method paper mills use to cut rolls
- The problem
- AI chat services keep a working memory for every conversation on the GPU. Conversations differ in length, so memory fragments and goes to waste, which limits how many users one GPU can serve.
- What to borrow
- Paper mills cut big rolls into orders of different widths with as little waste as possible. The classic method, column generation, copes with astronomically many cutting patterns and says how far a plan is from the best possible.
- What it gives
- Less wasted memory, more users per GPU, and a benchmark for judging the rules of thumb in use today.
Status. Not done. Current systems use rules of thumb and simple integer programs. Closest work: Online Scheduling for LLM Inference with KV Cache Constraints (2025) · Robust KV Cache Management for LLM Serving (2026)
Untouched 8Fields apart 4Real method 8Payoff 7
The other 83 pairings
24 more untouched ideas, ranked lower Nobody has done them, but a close cousin exists or the payoff is smaller.
- Complex systems + AI · ML · Data MiningStop errors growing along a chain of AI agents with the rule that keeps a convoy of self-driving trucks from amplifying each other's braking.Error growth is measured; this rule is not applied. 6.8
- Epidemics + SecurityEstimate how fast self-copying malicious prompts would spread among AI email assistants, and what share must be protected to stop them, like an epidemic's R number.Partly done. 6.8
- Religious studies + AI · ML · Data MiningTreat the family tree of open AI models, copied and fine-tuned by others, like cultural traits passed between people, and test what makes some spread.Family trees exist; the cultural-evolution test does not. 6.8
- Cognitive neuroscience + AI · ML · Data MiningUse the tools biologists use to order cells along a development path to trace how features form inside a language model during training.Features are tracked, but not with these tools. 6.5
- Music & Audio + Power & energyRecognise each appliance or solar inverter by its electrical 'voice', with the AI that recognises a speaker's voice.Similar methods exist. 6.5
- Religious studies + Network scienceTest whether people leave religion the way they take up a habit only after several friends have, using surveys that follow friendships over years.Simulated, not tested on real data. 6.5
- Time series + Power & energyForecast the swings in power use of AI training clusters with the volatility models used for stock prices, to size backup batteries.Not done, but a routine transfer. 6.5
- Dynamical systems + AI · ML · Data MiningPredict sudden failures when training large AI models from the warning signs that come before ecosystem tipping points: slower recovery, bigger wobbles.Other early-warning signals are already studied. 6.3
- Theoretical CS + PsychologyCompare how many glances people need to find an object with the theoretical minimum number of yes-or-no questions.Already compared with other ideal models. 6.3
- Control + AI · ML · Data MiningProve that a control system stays stable even when a language model that sometimes errs proposes its actions.Sketched in a survey; no proofs yet. 6.3
- Epidemics + SecurityPatch the devices next to a hacked one first, as ring vaccination does around a smallpox case, and compare with patching everyone in turn.Close variants exist. 6.3
- Visualization + TheologyMap which Bible verses each early Church Father quoted, and how the map shifted during the debates over the nature of Christ.Done for one author, Origen. 6.3
- Materials + BiomedicalModel a tumour's invading edge as a crack running through tissue, with the mathematics of fracture in materials.Related tumour models exist. 6.3
- Visualization + Sleep & dreamsMeasure how real lucid dreams feel with the questionnaires that measure 'being there' in virtual reality.Proposed in theory only. 6.0
- Materials + AI · ML · Data MiningDetect AI hallucinations with the tricks chemists use to tell when an atom-level simulation is out of its depth.Similar detectors already exist in AI. 6.0
- Power & energy + AI · ML · Data MiningPredict how one failing AI agent can bring down a team of agents, with the models used for blackouts spreading across a grid.Error cascades are studied with other models. 5.8
- Dynamical systems + AI · ML · Data MiningExplain why AI text generators get stuck repeating themselves with the mathematics of how steady systems start to oscillate.Other dynamical explanations exist. 5.8
- UQ & statistics + Sleep & dreamsTune sleep-deepening sound pulses to each person over a few nights with the search method used to tune AI models.Only a master's thesis so far. 5.8
- Visualization + MaterialsRebuild the 3D inside of a material from a few X-ray shots with the new graphics technique behind photo-real 3D scenes.Used for X-rays already, not tested on materials. 5.8
- Dynamical systems + AI · ML · Data MiningUse the mathematics of fireflies flashing in sync to predict when AI agents debating a question will agree or split.Related ideas exist; never fitted to real debates. 5.5
- Music & Audio + MaterialsPretrain a speech-style AI on hours of unlabelled crackling sounds from stressed metal and concrete, then use it to detect growing cracks.Smaller versions exist. 5.5
- Applied math + AI · ML · Data MiningLet an AI read long documents in detail only where it matters, as simulations refine their grid only where things happen.A close method appeared in 2026. 5.5
- Control + SecurityUse a safety filter from robotics to stop a dataset's privacy budget being overspent as queries arrive.Existing privacy filters already do most of this. 5.5
- Sleep & dreams + AI · ML · Data MiningTrain AI on new material in cycles that copy human sleep: more review early in the night, more 'dreaming' later.Sleep-like phases tried; not the human schedule. 5.5
4 already done The search found someone had got there first.
- Dynamical systems + AI · ML · Data MiningLook for tipping-point warning signs before an AI suddenly 'gets' a task after long training.Mostly done.
- Religious studies + AI · ML · Data MiningTest whether AI, like people, best remembers stories that are only slightly supernatural.Done: PNAS 2023 (Acerbi and Stubbersfield).
- Time series + AI · ML · Data MiningChoose AI models by how badly they fail in the worst cases, with financial risk measures, rather than by average score.Done: ICML 2024 (Nitsure et al.).
- Biomedical + TheologyUse medical-image AI to read erased text under rewritten medieval manuscripts.Mostly done with similar networks.
55 dropped early 21 already have papers joining the two fields; 34 were a loose analogy rather than a method that carries over.
- Cognitive neuroscience + AI · ML · Data MiningTrim neural networks with the rules brain immune cells use to prune unused connections.Already studied
- Applied math + AI · ML · Data MiningCompress a chatbot's working memory with the data-reduction method engineers use to speed up simulations.Already studied
- Physics + Complex systemsExplain jams in mixed self-driving and human traffic with the physics of self-propelled particles clumping.Already studied
- UQ & statistics + AI · ML · Data MiningCorrect AI leaderboards for the many 'A beats B' claims they make at once, so the lucky ones are not taken as real.Already studied
- Epidemics + AI · ML · Data MiningCorrect recent AI benchmark trends for results not yet published, as epidemiologists correct for late case reports.Already studied
- Psychology + AI · ML · Data MiningModel an AI's choice of word as evidence piling up to a threshold, the standard model of human decisions.Already studied
- Psychology + Sleep & dreamsTest whether sleep-deprived people misjudge when to trust an AI's advice.Already studied
- Religious studies + Complex systemsMeasure whether people who move in sync during rituals bond more, with the mathematics of synchronising oscillators.Already studied
- Complex systems + AI · ML · Data MiningModel reviewer effort and AI-written reviews at AI conferences as strategies competing over time.Already studied
- Network science + AI · ML · Data MiningCheck whether similar words are sent to the same 'expert' inside mixture-of-experts AI models, and what that does to the load balance.Already studied
- Time series + AI · ML · Data MiningTest whether reviewers' score changes after author rebuttals can be predicted, as markets are tested for efficiency.Already studied
- Time series + Sleep & dreamsPull a few common sleep patterns out of many nights of wearable data with the models economists use for many indicators.Already studied
- Optimization & OR + EpidemicsAllocate antiviral stockpiles so they hold up under the worst plausible epidemic curves.Already studied
- Materials + AI · ML · Data MiningDescribe how AI models lose the ability to learn new things with the laws for metal hardening as it is bent.Already studied
- Power & energy + AI · ML · Data MiningDesign conference reviewer incentives as a leader-and-follower optimisation.Already studied
- Theoretical CS + TheologyMeasure how tangled the arguments in Aquinas's Summa are with a graph measure from computer science.Already studied
- Biomedical + Religious studiesUse religious preference and chaplains' notes in medical records to predict end-of-life care choices.Already studied
- Security + Sleep & dreamsTrain sleep-scoring AI across clinics without sharing patient data, despite their different equipment.Already studied
- Security + Sleep & dreamsShow that encrypted traffic from bedroom smart devices can reveal sleep disorders.Already studied
- Cognitive neuroscience + AI · ML · Data MiningLook for the brain's daydreaming-network pattern in AI agents left with nothing to do.Already studied
- Cognitive neuroscience + Network scienceBuild a layered network linking gut microbes, the chemicals they make and brain regions.Already studied
- Dynamical systems + AI · ML · Data MiningCheck whether an AI stays stuck in a persona once pushed into it, a sign of hidden stable states.Analogy, not a method
- Optimization & OR + AI · ML · Data MiningSearch AI reasoning steps with a classic optimisation method that safely skips hopeless branches.Analogy, not a method
- Applied math + Network scienceModel how opinion groups online slowly merge with the equation for oil and water separating.Analogy, not a method
- Physics + AI · ML · Data MiningMeasure a geometric 'twist' that network weights pick up when the learning rate is cycled up and down.Analogy, not a method
- Hardware & EDA + AI · ML · Data MiningHave three AI agents vote on each action, as spacecraft computers do, and measure the gain in reliability.Analogy, not a method
- Hardware & EDA + Cognitive neuroscienceUse chip-design trade-off tools to ask whether brain wiring balances cost against performance.Analogy, not a method
- Epidemics + AI · ML · Data MiningEstimate whether AI coding assistants cause more or fewer bugs by imitating a randomised trial from project histories.Analogy, not a method
- Theology + AI · ML · Data MiningGrade the analogies AI models make with Thomas Aquinas's types of analogy.Analogy, not a method
- Theology + AI · ML · Data MiningLet AI read long documents by moving back and forth between the parts and the whole, as interpreters of texts do.Analogy, not a method
- Religious studies + SecurityKeep fake participants out of shared AI training by demanding costly contributions, as costly rituals prove religious commitment.Analogy, not a method
- Network science + Hardware & EDAWork out how many failures it takes to break the chip-design supply chain.Analogy, not a method
- Network science + AI · ML · Data MiningTest whether reviewer-paper assignments have the nested shape of ecosystems, which would make them robust to reviewers dropping out.Analogy, not a method
- Complex systems + AI · ML · Data MiningPredict which add-on modules for AI models combine well by looking at groups of them, not pairs.Analogy, not a method
- Complex systems + Power & energyPlan moving jobs out of a failing data centre with crowd-evacuation models.Analogy, not a method
- Complex systems + Hardware & EDAModel congestion in the wiring inside AI chips with the equations for highway traffic.Analogy, not a method
- Time series + TheologyFind the moments when theological vocabulary changed sharply across church councils.Analogy, not a method
- Optimization & OR + EpidemicsDecide when to start and end lockdowns with guarantees that hold however the epidemic unfolds.Analogy, not a method
- Optimization & OR + Sleep & dreamsSchedule nurses' shifts around each nurse's body-clock timing, measured by a wearable.Analogy, not a method
- Optimization & OR + Network sciencePick conference cities that minimise total flight emissions.Analogy, not a method
- Materials + AI · ML · Data MiningModel clashes when merging AI models with the energy at boundaries between grains in a metal.Analogy, not a method
- Power & energy + Sleep & dreamsTrack sleep pressure like a battery's charge, with the estimators in electric-car battery systems.Analogy, not a method
- Power & energy + Cognitive neuroscienceCompare how solar inverters keep the grid in rhythm with how the brain's thalamus sets its rhythms.Analogy, not a method
- Quantum computing + PsychologyCompare how well people detect very dim light with the quantum limit for the same light.Analogy, not a method
- Quantum computing + EpidemicsRoute vaccine deliveries that spoil over time with algorithms for routing fragile quantum links.Analogy, not a method
- Quantum computing + AI · ML · Data MiningModel how chatbots forget instructions in long chats as quantum states losing coherence.Analogy, not a method
- Theoretical CS + PhilosophyUse the length of the shortest proof as a measure of how costly a belief is to justify.Analogy, not a method
- Mathematics + Sleep & dreamsMeasure how jagged dream stories are compared with waking stories, using fractal dimension.Analogy, not a method
- Mathematics + AI · ML · Data MiningDescribe the ways to rearrange a network without changing what it computes with a branch of algebra, to help merge models.Analogy, not a method
- Mathematics + Cognitive neuroscienceFind bottlenecks in how brain cells develop with a geometric flow from mathematics.Analogy, not a method
- Applied math + TheologyDate the editing layers of biblical manuscripts as an inverse problem, like sharpening a blurred image.Analogy, not a method
- Applied math + Network scienceModel sudden surges of misinformation as shock waves travelling through a network.Analogy, not a method
- Security + PsychologyUse hidden-trigger attacks on AI as a model of traumatic memories set off by cues.Analogy, not a method
- Security + Religious studiesMeasure how privacy noise in census data distorts counts of small religious groups.Analogy, not a method
- Cognitive neuroscience + Dynamical systemsCompare the memory capacity of young and old brains' wiring when each is used as a simple computer.Analogy, not a method
How this was made
- I paired a method, tool or finding from one field with a concrete problem in another, 103 pairs in all, drawn from the 27 fields on the conference pages.
- For each pair I counted papers whose title or abstract names both sides, in Semantic Scholar and Europe PMC. Pairs that were already well studied, or where the link was only a loose analogy, stopped here.
- The 48 that remained were searched by hand, two to five searches each in different wordings, to find the closest existing work.
- Each surviving idea got four marks out of 10: how untouched it is, how far apart the two fields are, whether a real method carries over (not just a metaphor), and how much it would give. The score is their average. An idea can be untouched because it is pointless; the last two marks push those down.
Built 2026-10-24. Not finding a paper is evidence, not proof: someone may have done it under other words.