Staying on high of a fast-growing analysis subject is rarely simple.
I face this problem firsthand as a practitioner in Physics-Knowledgeable Neural Networks (PINNs). New papers, be they algorithmic developments or cutting-edge purposes, are revealed at an accelerating tempo by each academia and trade. Whereas it’s thrilling to see this speedy improvement, it inevitably raises a urgent query:
How can one keep knowledgeable with out spending numerous hours sifting by means of papers?
That is the place I’ve discovered overview papers to be exceptionally worthwhile. Good overview papers are efficient instruments that distill important insights and spotlight vital tendencies. They’re big-time savers guiding us by means of the flood of data.
On this weblog publish, I want to share with you my private, curated checklist of must-read overview papers on PINNs, which might be particularly influential for my very own understanding and use of PINNs. These papers cowl key points of PINNs, together with algorithmic developments, implementation greatest practices, and real-world purposes.
Along with what’s out there in present literature, I’ve included considered one of my very own overview papers, which gives a complete evaluation of widespread practical utilization patterns of PINNs — a sensible perspective typically lacking from tutorial evaluations. This evaluation is predicated on my overview of round 200 arXiv papers on PINNs throughout varied engineering domains up to now 3 years and might function a necessary information for practitioners trying to deploy these strategies to deal with real-world challenges.
For every overview paper, I’ll clarify why it deserves your consideration by explaining its distinctive perspective and indicating sensible takeaways you could profit from instantly.
Whether or not you’re simply getting began with PINNs, utilizing them to deal with real-world issues, or exploring new analysis instructions, I hope this assortment makes navigating the busy subject of PINN analysis simpler for you.
Let’s lower by means of the complexity collectively and give attention to what really issues.
1️⃣ Scientific Machine Studying by means of Physics-Knowledgeable Neural Networks: The place we’re and what’s subsequent
📄 Paper at a look
🔍 What it covers
- Authors: S. Cuomo, V. Schiano di Cola, F. Giampaolo, G. Rozza, M. Raissi, and F. Piccialli
- Yr: 2022
- Hyperlink: arXiv
This overview is structured round key themes in PINNs: the basic parts that outline their structure, theoretical points of their studying course of, and their software to numerous computing challenges in engineering. The paper additionally explores the out there toolsets, rising tendencies, and future instructions.
✨ What’s distinctive
This overview paper stands out within the following methods:
- Among the best introductions to PINN fundamentals. This paper takes a well-paced strategy to explaining PINNs from the bottom up. Part 2 systematically dissects the constructing blocks of a PINN, protecting varied underlying neural community architectures and their related traits, how PDE constraints are included, widespread coaching methodologies, and studying idea (convergence, error evaluation, and so forth.) of PINNs.
- Placing PINNs in historic context. Reasonably than merely presenting PINNs as a standalone answer, the paper traces their improvement from earlier work on utilizing deep studying to resolve differential equations. This historic framing is efficacious as a result of it helps demystify PINNs by displaying that they’re an evolution of earlier concepts, and it makes it simpler for practitioners to see what options can be found.
- Equation-driven group. As an alternative of simply classifying PINN analysis by scientific domains (e.g., geoscience, materials science, and so forth.) as many different evaluations do, this paper categorizes PINNs based mostly on the varieties of differential equations (e.g., diffusion issues, advection issues, and so forth.) they remedy. This equation-first perspective encourages information switch as the identical set of PDEs may very well be used throughout a number of scientific domains. As well as, it makes it simpler for practitioners to see the strengths and weaknesses of PINNs when coping with various kinds of differential equations.
🛠 Sensible goodies
Past its theoretical insights, this overview paper presents instantly helpful assets for practitioners:
- A whole implementation instance. In part 3.4, this paper walks by means of a full PINN implementation to resolve a 1D Nonlinear Schrödinger equation. It covers translating equations into PINN formulations, dealing with boundary and preliminary situations, defining neural community architectures, selecting coaching methods, deciding on collocation factors, and making use of optimization strategies. All implementation particulars are clearly documented for simple reproducibility. The paper compares PINN efficiency by various totally different hyperparameters, which might supply instantly relevant insights on your personal PINN experiments.
- Out there frameworks and software program instruments. Desk 3 compiles a complete checklist of main PINN toolkits, with detailed software descriptions supplied in part 4.3. The thought-about backends embody not solely Tensorflow and PyTorch but additionally Julia and Jax. This side-by-side comparability of various frameworks is very helpful for choosing the right software on your wants.
💡Who would profit
- This overview paper advantages anybody new to PINNs and searching for a transparent, structured introduction.
- Engineers and builders searching for sensible implementation steering would discover the real looking, hands-on demo, and the thorough comparability of present PINN frameworks most fascinating. Moreover, they’ll discover related prior work on differential equations much like their present drawback, which presents insights they’ll leverage in their very own problem-solving.
- Researchers investigating theoretical points of PINN convergence, optimization, or effectivity may significantly profit from this paper.
2️⃣ From PINNs to PIKANs: Latest Advances in Physics-Knowledgeable Machine Studying
📄 Paper at a look
- Authors: J. D. Toscano, V. Oommen, A. J. Varghese, Z. Zou, N. A. Daryakenari, C. Wu, and G. E. Karniadakis
- Yr: 2024
- Hyperlink: arXiv
🔍 What it covers
This paper gives one of many newest overviews of the newest developments in PINNs. It emphasises enhancements in community design, function growth, optimization methods, uncertainty quantification, and theoretical insights. The paper additionally surveys key purposes throughout a spread of domains.

✨ What’s distinctive
This overview paper stands out within the following methods:
- A structured taxonomy of algorithmic developments. Some of the recent contributions of this paper is its taxonomy of algorithmic developments. This new taxonomy scheme elegantly categorizes all of the developments into three core areas: (1) illustration mannequin, (2) dealing with governing equations, and (3) optimization course of. This construction gives a transparent framework for understanding each present developments and potential instructions for future analysis. As well as, the illustrations used within the paper are top-notch and simply digestible.

- Highlight on Physics-informed Kolmogorov–Arnold Networks (KAN). KAN, a brand new structure based mostly on the Kolmogorov–Arnold illustration theorem, is at the moment a sizzling matter in deep studying. Within the PINN group, some work has already been executed to interchange the multilayer perceptions (MLP) illustration with KANs to realize extra expressiveness and coaching effectivity. The group lacks a complete overview of this new line of analysis. This overview paper (part 3.1) precisely fills within the hole.
- Evaluate on uncertainty quantification (UQ) in PINNs. UQ is crucial for the dependable and reliable deployment of PINNs when tackling real-world engineering purposes. In part 5, this paper gives a devoted part on UQ, explaining the widespread sources of uncertainty in fixing differential equations with PINNs and reviewing methods for quantifying prediction confidence.
- Theoretical advances in PINN coaching dynamics. In follow, coaching PINNs is non-trivial. Practitioners are sometimes puzzled by why PINNs coaching generally fail, or how they need to be skilled optimally. In part 6.2, this paper gives one of the crucial detailed and up-to-date discussions on this side, protecting the Neural Tangent Kernel (NTK) evaluation of PINNs, data bottleneck idea, and multi-objective optimization challenges.
🛠 Sensible goodies
Regardless that this overview paper leans in direction of the theory-heavy aspect, two notably worthwhile points stand out from a sensible perspective:
- A timeline of algorithmic advances in PINNs. In Appendix A Desk, this paper tracks the milestones of key developments in PINNs, from the unique PINN formulation to the latest extensions to KANs. In case you’re engaged on algorithmic enhancements, this timeline offers you a transparent view of what’s already been executed. In case you’re scuffling with PINN coaching or accuracy, you should utilize this desk to search out present strategies that may remedy your difficulty.
- A broad overview of PINN purposes throughout domains. In comparison with all the opposite evaluations, this paper strives to provide the most complete and up to date protection of PINN purposes in not solely the engineering domains but additionally different less-covered fields resembling finance. Practitioners can simply discover prior works performed of their domains and draw inspiration.
💡Who would profit
- For practitioners working in safety-critical fields that want confidence intervals or reliability estimates on their PINN predictions, the dialogue on UQ could be helpful. If you’re scuffling with PINN coaching instability, gradual convergence, or sudden failures, the dialogue on PINN coaching dynamics may help unpack the theoretical causes behind these points.
- Researchers might discover this paper particularly fascinating due to the brand new taxonomy, which permits them to see patterns and determine gaps and alternatives for novel contributions. As well as, the overview of cutting-edge work on PI-KAN will also be inspiring.
3️⃣ Physics-Knowledgeable Neural Networks: An Utility-Centric Information
📄 Paper at a look
- Authors: S. Guo (this creator)
- Yr: 2024
- Hyperlink: Medium
🔍 What it covers
This text evaluations how PINNs are used to deal with various kinds of engineering duties. For every activity class, the article discusses the issue assertion, why PINNs are helpful, how PINNs could be carried out to deal with the issue, and is adopted by a concrete use case revealed within the literature.

✨ What’s distinctive
In contrast to most evaluations that categorize PINN purposes both based mostly on the kind of differential equations solved or particular engineering domains, this text picks an angle that practitioners care about probably the most: the engineering duties solved by PINNs. This work is predicated on reviewing papers on PINN case research scattered in varied engineering domains. The result is a listing of distilled recurring practical utilization patterns of PINNs:
- Predictive modeling and simulations, the place PINNs are leveraged for dynamical system forecasting, coupled system modeling, and surrogate modeling.
- Optimization, the place PINNs are generally employed to attain environment friendly design optimization, inverse design, mannequin predictive management, and optimized sensor placement.
- Knowledge-driven insights, the place PINNs are used to determine the unknown parameters or practical types of the system, in addition to to assimilate observational information to raised estimate the system states.
- Knowledge-driven enhancement, the place PINNs are used to reconstruct the sphere and improve the decision of the observational information.
- Monitoring, diagnostic, and well being evaluation, the place PINNs are leveraged to behave as digital sensors, anomaly detectors, well being screens, and predictive maintainers.
🛠 Sensible goodies
This text locations practitioners’ wants on the forefront. Whereas most present overview papers merely reply the query, “Has PINN been utilized in my subject?”, practitioners typically search extra particular steering: “Has PINN been used for the kind of drawback I’m attempting to resolve?”. That is exactly what this text tries to deal with.
Through the use of the proposed five-category practical classification, practitioners can conveniently map their issues to those classes, see how others have solved them, and what labored and what didn’t. As an alternative of reinventing the wheel, practitioners can leverage established use instances and adapt confirmed options to their very own issues.
💡Who would profit
This overview is greatest for practitioners who wish to see how PINNs are literally being utilized in the actual world. It will also be notably worthwhile for cross-disciplinary innovation, as practitioners can study from options developed in different fields.
4️⃣ An Professional’s Information to Coaching Physics-informed Neural Networks
📄 Paper at a look
- Authors: S. Wang, S. Sankaran, H. Wang, P. Perdikaris
- Yr: 2023
- Hyperlink: arXiv
🔍 What it covers
Regardless that it doesn’t market itself as a “commonplace” overview, this paper goes all in on offering a complete handbook for coaching PINNs. It presents an in depth set of greatest practices for coaching physics-informed neural networks (PINNs), addressing points like spectral bias, unbalanced loss phrases, and causality violations. It additionally introduces difficult benchmarks and intensive ablation research to exhibit these strategies.

✨ What’s distinctive
- A unified “professional’s information”. The primary authors are lively researchers in PINNs, working extensively on bettering PINN coaching effectivity and mannequin accuracy for the previous years. This paper is a distilled abstract of the authors’ previous work, synthesizing a broad vary of current PINN strategies (e.g., Fourier function embeddings, adaptive loss weighting, causal coaching) right into a cohesive coaching pipeline. This looks like having a mentor who tells you precisely what does and doesn’t work with PINNs.
- An intensive hyperparameter tuning research. This paper conducts varied experiments to indicate how totally different tweaks (e.g., totally different architectures, coaching schemes, and so forth.) play out on totally different PDE duties. Their ablation research present exactly which strategies transfer the needle, and by how a lot.
- PDE benchmarks. The paper compiles a set of difficult PDE benchmarks and presents state-of-the-art outcomes that PINNs can obtain.
🛠 Sensible goodies
- An issue-solution cheat sheet. This paper totally paperwork varied strategies addressing widespread PINN coaching pain-points. Every method is clearly offered utilizing a structured format: the why (motivation), how (how the strategy addresses the issue), and what (the implementation particulars). This makes it very simple for practitioners to determine the “remedy” based mostly on the “signs” noticed of their PINN coaching course of. What’s nice is that the authors transparently mentioned potential pitfalls of every strategy, permitting practitioners to make well-informed choices and efficient trade-offs.
- Empirical insights. The paper shares worthwhile empirical insights obtained from intensive hyperparameter tuning experiments. It presents sensible steering on selecting appropriate hyperparameters, e.g., community architectures and studying charge schedules, and demonstrates how these parameters work together with the superior PINN coaching strategies proposed.
- Prepared-to-use library. The paper is accompanied by an optimized JAX library that practitioners can instantly undertake or customise. The library helps multi-GPU environments and is prepared for scaling to large-scale issues.
💡Who would profit
- Practitioners who’re scuffling with unstable or gradual PINN coaching can discover many sensible methods to repair widespread pathologies. They will additionally profit from the easy templates (in JAX) to shortly adapt PINNs to their very own PDE setups.
- Researchers searching for difficult benchmark issues and aiming to benchmark new PINN concepts in opposition to well-documented baselines will discover this paper particularly useful.
5️⃣ Area-Particular Evaluate Papers
Past common evaluations in PINNs, there are a number of good overview papers that target particular scientific and engineering domains. In case you’re working in considered one of these fields, these evaluations might present a deeper dive into greatest practices and cutting-edge purposes.
1. Warmth Switch Issues
Paper: Physics-Informed Neural Networks for Heat Transfer Problems
The paper gives an application-centric dialogue on how PINNs can be utilized to deal with varied thermal engineering issues, together with inverse warmth switch, convection-dominated flows, and phase-change modeling. It highlights real-world challenges resembling lacking boundary situations, sensor-driven inverse issues, and adaptive cooling system design. The commercial case research associated to energy electronics is especially insightful for understanding the utilization of PINNs in follow.
2. Energy Techniques
Paper: Applications of Physics-Informed Neural Networks in Power Systems — A Review
This paper presents a structured overview of how PINNs are utilized to vital energy grid challenges, together with state/parameter estimation, dynamic evaluation, energy circulation calculation, optimum energy circulation (OPF), anomaly detection, and mannequin synthesis. For every kind of software, the paper discusses the shortcomings of conventional energy system options and explains why PINNs may very well be advantageous in addressing these shortcomings. This comparative abstract is helpful for understanding the motivation for adopting PINNs.
3. Fluid Mechanics
Paper: Physics-informed neural networks (PINNs) for fluid mechanics: A review
This paper explored three detailed case research that exhibit PINNs software in fluid dynamics: (1) 3D wake circulation reconstruction utilizing sparse 2D velocity information, (2) inverse issues in compressible circulation (e.g., shock wave prediction with minimal boundary information), and (3) biomedical circulation modeling, the place PINNs infer thrombus materials properties from phase-field information. The paper highlights how PINNs overcome limitations in conventional CFD, e.g., mesh dependency, costly information assimilation, and problem dealing with ill-posed inverse issues.
4. Additive Manufacturing
Paper: A review on physics-informed machine learning for monitoring metal additive manufacturing process
This paper examines how PINNs deal with vital challenges particular to additive manufacturing course of prediction or monitoring, together with temperature subject prediction, fluid dynamics modeling, fatigue life estimation, accelerated finite ingredient simulations, and course of traits prediction.
6️⃣ Conclusion
On this weblog publish, we went by means of a curated checklist of overview papers on PINNs, protecting basic theoretical insights, the newest algorithmic developments, and sensible application-oriented views. For every paper, we highlighted distinctive contributions, key takeaways, and the viewers that may profit probably the most from these insights. I hope this curated assortment may help you higher navigate the evolving subject of PINNs.