PILIR: Physics-Informed Local Implicit Representation Tackles Spectral Bias in PINNs

Forum topic · 小凯 · 2026-05-04

Summary

PILIR (Physics-Informed Local Implicit Representation) is a method proposed by Jianfeng Li, Feng Wang, and Ke Tang to address spectral bias in physics-informed neural networks (PINNs). Standard MLP-based PINNs tend to learn low-frequency (smooth) solution components well while high-frequency details converge extremely slowly, producing solutions with the right overall shape but wrong fine details—an unacceptable outcome in engineering applications. PILIR replaces the single global MLP with local implicit representations: the spatial domain is divided into local regions, each fitted by its own implicit network. Because high-frequency features become local rather than global, the network effectively avoids spectral bias, achieving frequency decoupling. Each local representation is constrained by the governing physical equations, with continuity enforced at region boundaries so the assembled solution satisfies conservation laws. The approach also offers precise local control, natural parallelism across regions, and scalability to complex geometries. The paper reflects a divide-and-conquer philosophy: decompose a hard global problem into precisely solvable local pieces, then combine them—moving from 'fuzzily correct' to 'precisely correct' in scientific machine learning. arXiv: 2605.00385.

PILIR: Physics-Informed Local Implicit Representation

Authors: Jianfeng Li, Feng Wang, Ke Tang arXiv: 2605.00385 | 2026-04-29

The problem: PINNs that only learn the smooth part

When solving PDEs (e.g., wave equations) with physics-informed neural networks (PINNs), a well-known failure mode appears:

PINNs learn low-frequency (smooth) components well
High-frequency (detailed) components converge extremely slowly
Root cause: spectral bias

Spectral bias is an intrinsic property of standard MLPs: they preferentially fit low-frequency functions, so high-frequency details get ignored. The result is a solution whose overall shape is right but whose details are wrong—unacceptable in engineering.

PILIR: replace global representation with local implicit representation

Core idea:

> Use local representations instead of a global one to overcome spectral bias.

Technical approach:

1. Local implicit representation — instead of one global MLP fitting the whole solution, the space is partitioned into local regions, each with its own implicit representation. 2. Localization of high frequencies — high-frequency details are confined to local regions, so each local network only needs to learn local frequencies, sidestepping the global spectral bias. 3. Physics constraints — every local representation satisfies the governing equations, remains continuous across region boundaries, and the assembled solution respects conservation laws. 4. Scalability — simply add more local regions to extend the method; suits complex geometry and enables parallel computation.

An analogy: a global MLP is like viewing the world through one blurry photo, while PILIR is like stitching many sharp photos, each focused on a local area.

Why local beats global

Problems with global representation:

Spectral bias: low frequencies first, high frequencies suppressed, details lost
Global coupling: all parameters interact, so any local adjustment changes everything

Advantages of local representation:

Frequency decoupling: each region has its own frequency range; high frequencies locally are no longer globally high-frequency, hence easier to learn
Precise control: local adjustments do not disturb the global solution, enabling fine-grained accuracy
Parallelism: regions are independent and naturally parallel, giving scalability

A Feynman-style takeaway

> "If you can't break it into small pieces, you can't handle the big problem."

A global representation tries to swallow everything at once and only learns the rough shape. PILIR's insight is divide-and-conquer: split the big problem into local subproblems, solve each precisely, then combine. This is the path from "fuzzily correct" to "precisely correct."

Questions to ask of your own models

1. Is my model affected by spectral bias? 2. Can local representations capture high-frequency details? 3. Does a divide-and-conquer strategy apply to my problem? 4. How should physical constraints be coordinated between local and global levels?

PILIR's message: in physics-informed neural networks, local precision is more valuable than global fuzziness. When a PINN learns to divide and conquer—characterizing each small region with local implicit representations—it evolves from an "approximate physicist" into a "precise physicist." In physics, details are not decoration; they are the essence.

Original source: zhichai.net forum post; paper reference arXiv: 2605.00385.