Research

Research programmes and
mathematical architecture.

The age of AI is changing the scale of scientific problems that can be explored. I have chosen two principal research programmes in response: stable reduction of unresolved dynamics in large, multiscale systems, and mathematically certified numerical computation. They address distinct questions, but both seek to make increasingly ambitious computation mathematically meaningful.

AIの発展によって,科学研究で扱える問題の規模や複雑さは大きく変わろうとしています.私はこの時代を見据え,大規模・マルチスケール系における未解像ダイナミクスの安定縮約と,有限要素法を中心とする数値計算の数学的保証を研究の二つの柱に据えています.より多くの変数や複雑な現象に挑戦することと,得られた計算結果をどこまで信頼できるのか数学的に明らかにすること.この両方を進めたいと考えています.

Core programmes

Two distinct theories.

Two mathematically distinct programmes, chosen for an era in which AI-assisted research can explore ever larger and more complex systems, while mathematical reliability remains essential.

AIによって大きな問題を探索できる可能性が広がる一方で,計算結果の信頼性を数学的に確かめる必要があります.この二つの研究は,その問題に異なる方向から取り組むものです.

Programme I

Mathematical foundations of certified finite element analysis

I investigate when finite element approximations remain mathematically reliable. Starting from computable geometric conditions and anisotropic interpolation, I study how approximation spaces, stability and the preservation of mathematical structure affect the properties that discretisation can retain and the errors it can control.

異方性メッシュの幾何条件や補間理論を出発点に,有限要素離散化が安定性,保存則,単調性などをどの条件で保持できるのかを考えます.さらに,数値解の正しさをどこまで計算可能な誤差評価で保証できるかを追究します.

Computable geometryAnisotropic FEMCR / RT / MorleyStructure preservationPressure robustnessError certification
Established foundations →
Programme II

Stable reduction of unresolved dynamics

I investigate what information is lost when unresolved degrees of freedom are eliminated, which effects must be retained, and when the resulting dynamics can be trusted. Memory and delay provide mathematical representations of those effects, but their presence alone does not establish a stable reduction.

未解像自由度を消去したとき,何の情報が失われ,どの影響を残さなければならないのかを問います.記憶や遅れはその解析手段です.最終的には,縮約を許容できる場合,許容できない場合,現時点の情報では判定できない場合を数学的に区別する理論を目指します.

Unresolved dynamicsInformation lossReduction criteriaStability boundariesMemory and delayCertified approximation
Research questions →

Why these two programmes? — なぜこの二つの研究テーマなのか

I have deliberately chosen problems whose scale and complexity make them difficult to explore by hand. AI-assisted computation may allow us to investigate systems with far more coupled variables, widely separated time scales and richer parameter dependence than before. I want the mathematics to grow alongside that capacity, rather than simply use AI to accelerate familiar calculations.

Unresolved-dynamics theory asks which effects of discarded degrees of freedom must be retained to describe a large system faithfully. Certified finite element analysis, building on my work in anisotropic numerical analysis, asks how to control approximation error and assess the reliability of computed solutions. The first direction confronts the scale of the problem; the second confronts the credibility of its computational answer. Both programmes have their own mathematical questions, but together they respond to the same challenge of AI-era scientific computation.

AI時代を見据えた三つの研究方針

  1. AI時代だからこそ,大規模で複雑な数学的問題に挑戦する. 人間だけでは探索が難しかった高次元系,マルチスケール問題,未解像自由度を含む力学系に取り組みます.
  2. AIの計算能力が向上するほど,研究の可能性も広がるテーマを選ぶ. 単に既存の問題を速く解くのではなく,より大きな系や複雑な構造を扱える研究体系を目指します.
  3. 計算能力がどれほど向上しても,数学的保証は不可欠である. AIを利用して得られた予測,数値解,縮約モデルの信頼性を,誤差評価・安定性・保存構造などによって明らかにします.

これらは,二つの研究テーマを選んだ理由でもあります.未解像ダイナミクスでは,多数の自由度を持つ系において,何を消去し,何を残せば元の現象を記述できるかを問います.数学的保証では,近似・離散化・縮約に伴う誤差や構造の変化を評価し,どこまで計算結果を信頼できるかを問います.一方では問題の大きさに挑戦し,他方では計算結果の信頼性を追究する.それぞれ独立した数学的課題を持ちながら,AI時代の科学計算を支える二つの方向だと考えています.

How I envisage human-guided mathematical discovery with AI →

FEM 2.0

Five goals.

Five complementary aims organise the mathematical analysis, construction and verification of reliable numerical methods.

EfficiencyApproximate complex phenomena with the resolution that is actually needed.
AbstractionExpose the mathematical structures shared by schemes, estimates and geometric conditions.
AutomationUse computable indicators to guide meshes, methods and adaptive numerical choices.
ApplicationCarry the theory into anisotropic PDEs, flow, history-dependent equations and complex geometry.
VerificationTurn numerical output into a mathematically interpretable and checkable certificate.
Research lines

G / R / A / V

The four G / R / A / V lines are cross-cutting research directions, not additional programmes. Individual projects in Programme I or II may draw on several lines.

G

Geometry & Anisotropic FEM Theory

Computable geometric parameters, semi-regular meshes, anisotropic interpolation, nonconforming and mixed elements, discrete functional inequalities and curved-domain geometry.

R

Reliable PDE Simulation & Reduction

Certified and goal-oriented simulation, time- and history-dependent PDEs, pressure-robust flow methods, graph-space analysis and stable unresolved-dynamics reduction.

A

Approximation, Automation & Learning

Adaptive FEM, mesh and method selection, data-informed kernel identification, structure-preserving approximation and comparisons with PINNs, Deep Ritz and meshless methods.

V

Verification, Formal Reasoning & AI4Math

Computable certificates, equilibrated fluxes, formalisation-ready assumptions and estimates, AI-assisted reasoning and reproducible mathematical structure.

AI4Math · Mathematical Assurance

Mathematics for trustworthy scientific AI

The rapid development of AI is part of my motivation for choosing research problems at a larger scale. I want AI to expand the range of systems and phenomena we can explore, while mathematics supplies the stability analysis, error bounds and structural understanding needed to interpret the results. This is why I pursue both unresolved-dynamics reduction and certified numerical analysis, rather than viewing AI4Math solely as automated theorem proving.

learning / reduction / discretisation→information loss→mathematical certificate→reliable scientific prediction

The long-term goal is to connect certified PDE computation, stable reduction of unresolved dynamics, structure-preserving learning and computable error control into scientific AI systems that can assess and improve the reliability of their own numerical predictions.

AI4MathMathematical assuranceCertified PDE computationStable reductionStructure-preserving learningSelf-correcting computation

Guiding principle: not only AI for mathematics, but mathematics for trustworthy AI.

Research Vision 17Human-guided exploration →
AI4Math · Human-guided discovery

From large-scale exploration to mathematical guarantees.

A research methodology for unresolved dynamics and multiscale systems, where numerical exploration is directed by mathematical questions and followed by independent analysis.

I am exploring a research cycle in which AI helps design and run structured computational experiments, organise the resulting data and search for counterexamples. Human mathematical judgement then identifies meaningful phenomena, formulates hypotheses and directs the next stage of analysis. The cycle is intended to support mathematical discovery, not to replace proof or verification.

human: formulate the question→AI-assisted computation→human: interpret the data→proof and certification→next question
01

Formulate

Specify the PDE or dynamical system, the resolved observables, assumptions, norms and quantities to be tested.

02

Explore

Use AI-assisted coding and analysis to vary unresolved dimensions, scale separation, coefficients and initial data. Record reproducible outputs.

03

Interpret

Examine trends and failures critically; identify candidate invariants, thresholds, memory effects and counterexamples.

04

Certify

Separate model-reduction and discretisation errors, audit arguments, and seek rigorous bounds or conditions that explain the observations.

An initial test case is a coupled resolved–unresolved linear evolution system. Eliminating the unresolved variables produces a memory term exactly; numerical experiments can then probe which stability and error estimates remain uniform as the unresolved dimension or time-scale separation changes. Observed patterns are conjectures, not established theorems, until the mathematical assumptions and proof have been verified.

This approach is designed to benefit from advances in computational capability while retaining the central responsibility of mathematical research: deciding what a result means, what has actually been established and what should be studied next. It complements the Education & Learning principle of learning to verify and formulate one's own questions.

Read the research-methodology essay → Unresolved dynamics programme →
Published foundations · Programme I

Results established in finite element analysis.

These topics are supported by peer-reviewed work. Ongoing certification projects and proposed extensions are listed separately below.

01

Computable geometric conditions

Geometric flatness parameters and their relation to classical maximum-angle conditions for triangles and tetrahedra.

02

Anisotropic interpolation theory

Direction-sensitive interpolation error bounds, including Lagrange and Raviart–Thomas estimates under explicit geometric assumptions.

03

CR and RT finite element analysis

Nonconforming and mixed methods, consistency estimates and CR–RT relations on anisotropic meshes.

04

Morley methods for fourth-order equations

Anisotropic interpolation and consistency analysis of nonconforming Morley finite elements under a semi-regular mesh condition.

05

Nitsche and discontinuous Galerkin methods

Weak boundary conditions, anisotropic WOPSIP methods and hybrid formulations with stability and error estimates.

06

Anisotropic Stokes analysis

Stability, discrete inf-sup estimates and error analysis for Stokes discretisations on anisotropic meshes.

07

Pressure-robust Navier–Stokes discretisation

Modified Crouzeix–Raviart methods using Raviart–Thomas-based velocity reconstruction for stationary incompressible flow.

08

Discrete Sobolev inequalities

Mesh-robust discrete Sobolev estimates for nonconforming finite elements under semi-regular mesh conditions.

See peer-reviewed publications →

Current work and open questions

Directions under investigation.

These are research objectives, not claims that a complete certification or reduction theory has already been established.

I · Certification

Reconstruction and rigorous error bounds

CR–RT equilibrated fluxes, hypercircle-type bounds and goal-oriented error control, with computable assumptions and explicit error contributions.

I · Pointwise structure

Monotonicity and maximum-norm guarantees

Discrete maximum principles, local Z-matrix criteria, inverse monotonicity and the limits imposed by anisotropic geometry.

I · Geometry

Curved and evolving domains

Extending geometry-aware estimates and finite element constructions to curved elements and time-dependent geometry while tracking geometric errors.

II · Reduction

Criteria for stable reduction

Quantifying the influence of eliminated degrees of freedom and seeking conditions under which reduction can be accepted, rejected or left uncertified.

Research architecture · Programme I

From geometric analysis to mathematical assurance.

The established analytical foundation connects element geometry, affine and Piola transformations, interpolation, stability, consistency and PDE error estimates. The same framework includes standard isotropic meshes while making anisotropic geometric deterioration explicit. Current research asks how this foundation can be extended to structural properties and computable certificates.

Established analytical foundation

Geometry, interpolation, stability, consistency and PDE error estimates

Computable geometric information enters error analysis through affine and Piola mappings, directional interpolation bounds and estimates for nonconforming, mixed and discontinuous schemes.

Geometry→Transformation→Interpolation→Stability & consistency→PDE error estimates
Further questions

Beyond convergence rates, I investigate which discrete structures can be retained and which additional reconstruction or verification steps make approximation errors mathematically assessable. This is a research direction, not a single established implication.

Discrete structure→Conservation / monotonicity→Reconstruction→Computable certification

The research note [N2] Interpolation error analysis using a new geometric parameter supplies the foundation for geometry, transformations and interpolation. Published CR, RT, Morley, WOPSIP, Nitsche and incompressible-flow analyses provide subsequent results; curved-domain extensions and rigorous output certification remain active lines of investigation.

Programme II · Open mathematical questions

Stable reduction theory for unresolved dynamics

Eliminating unresolved variables from evolution equations need not remove their dynamical influence. Memory, delay, history forcing and internal-state representations can describe retained effects, but such representations alone do not guarantee stability or fidelity to the full system.

Information lossReduction criteriaPositive-type memoryGraph-space well-posednessMeasure-valued delayKernel stability

The proposed research route is to quantify the influence of eliminated states and develop admissibility criteria; the final certification step is an aim rather than an established theorem.

Full dynamics→Eliminate unresolved states→Quantify information loss→Test admissibility→Certify if justified

I seek criteria that distinguish admissible, inadmissible and presently uncertifiable reductions. Current mathematical tools include operator-valued memory, coercivity and stability bounds. Further questions concern the derivation of closure operators, full-versus-reduced energy structures, and whether reduction and discretisation commute.

Research Vision 18Read the Blog entry route
Detailed research map

Five goals × four research lines.

The matrix connects the five long-term goals with the four cross-cutting research lines. Its cells describe directions for investigation, not a list of completed theorems.

Research line \ GoalEfficiencyAbstractionAutomationApplicationVerification
G · Geometry & anisotropic FEMReduce cost by directional interpolation and refine only where geometry or solution structure requires it.Organise flatness, semi-regularity, discrete Sobolev inequalities and related mesh conditions into computable geometry.Design anisotropic AFEM driven by geometric parameters.Extend FEM geometry to spheres, polygonal domains, CutFEM and other complex geometries.Make mesh assumptions explicit and checkable through computable geometric conditions.
R · Reliable & goal-oriented PDE simulationTarget the dominant seminorm error or quantity of interest rather than over-resolving every component.Abstract goal functionals, dual problems and goal-oriented error-control structures.Combine functional / hypercircle majorants, CR–RT fluxes and verified adaptive computation.Pressure-robust Stokes/Navier–Stokes, fractional-time PDEs and long-time stable simulations.Provide certified outputs through equilibrated fluxes, DWR and a posteriori certificates.
A · Automation, learning & new methodsAutomatically tune parameters and mesh strategies from mathematical and data-driven indicators.Compare FEM, PINNs and Deep Ritz through the common ideas of representation, metric and projection.Use geometric parameters and error indicators as features for mesh generation and method selection.Connect numerical analysis with climate, disaster, engineering and data-science applications.Add a reliability layer to physics-informed ML and surrogate models.
V · Verification, formal reasoning & AI4MathUse certified bounds for stopping criteria and efficient adaptation.Restructure interpolation, stability, conservation and error bounds into formalisation-ready statements.Support lemma search, proof assistance and counterexample detection in FEM analysis.Present certified outputs suitable for digital twins and decision-support workflows.Build a finite-element foundation that remains verifiable in an AI4Math setting.
Selected open problems

Mathematical questions for further study.

Specific questions that extend the present programme beyond established results.

DG medius analysis under semi-regular geometry

  • Identify what anisotropic elements change in standard DG analysis.
  • Clarify penalty terms, condition numbers and error estimates.
  • Develop a precise programme of assumptions, estimates and counterexamples.

Helmholtz decomposition on non-simple domains

  • Multiply connected domains and domains with corners.
  • Connection with pressure robustness.
  • Influence of boundary regularity and topology.

Lagrange FEM and hypercircle certification

  • Transfer anisotropic information through RT-type flux estimates.
  • Investigate whether equilibrated fluxes bypass low-regularity interpolation restrictions in 3D.
  • Determine how computable geometric assumptions enter certified computation.
Read full detailed research archive (Japanese)

Earlier detailed research notes are kept here as an archive. Their priorities and status labels may not reflect the current programme.

Load on open.
Research memo

Near-term mathematical priorities.

The discrete Sobolev foundations are part of the published programme. The following topics identify additional guarantees still being investigated.

Monotonicity and maximum-norm analysis

  • Geometric conditions for local Z-matrix structure and discrete maximum principles
  • Relations between local sign conditions and global inverse monotonicity
  • Limits and counterexamples on anisotropic meshes

Certified finite element output

  • CR–RT equilibrated flux and hypercircle-type bounds
  • Rigorous quantity-of-interest estimation and DWR comparisons
  • Computable constants, explicit assumptions and verification-ready statements