KANDy points to an important direction for AI: moving from models that predict outcomes well to models that attempt to discover the underlying rules that produced them. If the approach proves effective outside research settings, it could help enterprises understand complex physical systems, improve operational control, and shorten research and development cycles.
The distinction is important from the outset. KANDy is a research framework published as a preprint, not a production-ready enterprise product. Academic attention, open-source code, and promising results on difficult problems justify serious evaluation, not declarations of a proven solution.
Key insight: Prediction tells us what may happen. Equation discovery attempts to explain which forces and relationships cause it to happen. That distinction matters in research, engineering, and risk management.
What KANDy Actually Does
KANDy stands for Kolmogorov-Arnold Networks for Dynamics. The framework is based on the KAN architecture, a family of neural networks in which learned functions play a central role in representing relationships between variables.
A conventional predictive model takes a system's history and estimates its future state. KANDy tries to go one step further: it takes time-series data from a dynamic system and attempts to reconstruct the mathematical structure governing how that system changes over time.
In abstract terms, rather than learning only the transition from a state x(t) to a future state, the objective is to approximate the governing law dx/dt = f(x). The desired output is not merely a forecast. It is a representation that experts can inspect, compare with physical knowledge, and challenge professionally.
Researchers at Clarkson University tested the framework on continuous and discrete dynamic systems, chaotic partial differential equations, and abstract mathematical structures including the Hopf fibration. According to the study's results, KANDy reconstructed structures that comparison methods struggled to discover.
Interpretability is not a design feature to add at the end of a project. In critical systems, it is part of the definition of quality alongside accuracy, stability, and safety.
Why a Black Box Is Not Enough
A model can predict a failure with useful accuracy and still fail to answer the more important question: did it learn a causal relationship, a physical law, or merely a temporary correlation created by the measurement conditions?
That difference is critical in chaotic systems. Sensitivity to initial conditions means that a small measurement change can develop into a large forecasting error. A model that fits historical data well may look impressive in a standard evaluation, yet break down when operating conditions, equipment, or the production environment changes.
The main approaches can be compared as follows:
- Predictive model: Produces a future forecast and can deliver fast performance, but may depend heavily on correlations.
- Simulation model: Describes behavior according to known laws and gives domain experts greater control, but requires an existing model.
- Equation discovery: Infers governing relationships from data, supporting interpretation and discovery, but is sensitive to data quality.
- Hybrid model: Combines established laws with machine learning, balancing prior knowledge and learning at the cost of greater engineering complexity.
KANDy is particularly interesting because it sits between scientific discovery and machine learning engineering. It does not eliminate simulations, domain experts, or physical models. Its value may come precisely from connecting them: generating a mathematical hypothesis from data, testing it against established knowledge, and integrating it into a control system only after validation.
Potential Business Applications
This is not an immediate solution for every department with a spreadsheet. The approach is most relevant to enterprises whose core processes depend on measurable dynamic systems, including manufacturing, energy, water, transportation, climate, medical devices, biotechnology, and defense industries.
Potential applications include:
- Predictive maintenance: Identifying relationships that explain how vibration, heat, pressure, or wear develops, rather than only estimating when a failure will occur.
- Process control: Discovering relationships among production variables to design more stable control policies.
- Digital twins: Completing a physical model when part of the system's dynamics is unknown or too expensive to calculate.
- Research and development: Producing candidate equations that researchers can test experimentally.
- Energy management: Understanding interactions among load, temperature, equipment, and environmental conditions.
- Anomaly detection: Distinguishing measurement noise from a genuine change in system dynamics.
The financial value does not come from discovering an equation by itself. It appears when the new relationship changes an operational decision by reducing downtime, limiting the number of experiments, improving throughput, saving energy, or preventing unsafe operation.
How Enterprises Should Evaluate KANDy
A serious proof of concept should begin with a domain question, not a model choice. The enterprise should select a system whose behavior has economic significance, for which sufficient measurement history exists, and where internal experts can judge whether the result is plausible.
- Define the decision: Select an engineering or operational decision that could improve through a better understanding of the system's dynamics.
- Audit the data: Examine calibration, sampling frequency, missing variables, and changes in operating conditions.
- Discover candidates: Run the framework to produce possible mathematical structures, not a single unquestionable truth.
- Test out of sample: Evaluate operating regimes, disturbances, and edge conditions that did not appear during training.
- Validate with experts: Ask domain specialists to examine units, symmetries, stability, and physical plausibility.
- Connect it to the process: Deploy only after defining monitoring, exception handling, accountability, and rollback mechanisms.
A sound evaluation cannot rely on average error alone. It should ask whether the equation preserves physical units, remains stable across operating ranges, produces consistent parameters across experiments, and can be falsified through a designed experiment.
The comparison must also be fair. KANDy should be tested against an existing physical model, symbolic regression methods, state-space models, and a conventional predictive network. In some cases, a simpler model will provide greater business value at a lower cost. Novelty is not an exemption from systems engineering.
Keeping Humans in the Loop Without Creating a Bottleneck
Automated discovery of governing relationships is a strong example of a nondeterministic process in which human judgment remains essential. A model may produce a candidate that is mathematically persuasive but scientifically wrong, particularly when the data is incomplete or contains an unmeasured confounding variable.
Yet if an expert must manually approve every calculation and prediction, the organization has achieved no operational step change. A well-designed process divides responsibility:
- The system generates candidates, checks constraints, and ranks confidence.
- Software controls reject results that violate units, permitted ranges, or safety rules.
- An expert reviews new, unusual, or high-impact patterns.
- Approved models are stored in a managed repository with their version, data provenance, and validity period.
- Continuous monitoring detects drift and sends people only the events that require a decision.
The goal is to let an expert who previously supervised one experiment or process oversee dozens or even hundreds of runs, depending on the level of risk. Human-in-the-loop design should extend human capacity, not reproduce manual work inside a new interface.
This Is a Multidisciplinary Field, Not a Routine Software Project
KANDy illustrates why AI is not purely a technical matter. A team may know how to run a Python library yet lack an understanding of dynamic systems, statistics, measurement noise, or the business context. Such a team can produce an elegant but meaningless equation.
The work requires several forms of expertise:
- Domain researchers and engineers who understand the system and its physical constraints.
- Machine learning practitioners who understand optimization, generalization, uncertainty, and data bias.
- Data and MLOps professionals who provide traceability, versioning, monitoring, and experiment reproducibility.
- Executives and operational leaders who connect the research to decisions, costs, risks, and accountability.
- Safety, legal, and cybersecurity specialists when the result affects a critical system.
Academic research has a central role here. Work at the intersection of mathematics, physics, engineering, and AI produces knowledge that cannot be replaced by superficial familiarity with off-the-shelf tools. Researchers with combined expertise may have a significant advantage because they can formulate a scientific question and determine whether the proposed solution is practical.
Watch out: Not every equation is a discovery. A model may fit a formula to noise or a temporary correlation. Operational use requires external validation, sensitivity testing, and review by domain experts.
The risk is especially high for small and medium-sized businesses that rely on self-described AI experts. A project like this is not suitable for opportunistic consulting that promises results after a short demonstration. Buyers should require relevant education, applied experience, business understanding, and a clear validation methodology.
What KANDy Has Not Yet Proved
Results across a collection of research problems do not establish robustness on enterprise data. In a factory or laboratory, data is affected by replaced sensors, maintenance work, operators, environmental conditions, and operational policies. Some variables are not measured at all, while others are sampled at frequencies that do not match the underlying dynamics.
Publication as a preprint allows the community to examine the work early, but it is not equivalent to full peer review. Source-code availability is an important advantage for reproduction and evaluation, yet open source does not guarantee data quality, numerical stability, or production readiness.
An enterprise evaluating the framework should document at least:
- Which data and code versions were used in each experiment.
- How variables, ranges, and hyperparameters were selected.
- The conditions under which the equation is valid and those under which it is not.
- The level of uncertainty and the triggers for escalation to a person.
- Who has authority to approve a model for operational use.
- How the organization will return to an earlier version if performance deteriorates.
The Strategic Meaning Extends Beyond KANDy
The important story is not limited to one research tool. KANDy represents a possible shift from AI that imitates human output to AI that helps formulate new governing relationships. For knowledge-intensive enterprises, that could change how work is divided among researchers, engineers, and computational systems.
The organizations that benefit will not necessarily be those that adopt the library first. They will be the ones that have built scientific data infrastructure, internal machine learning and MLOps capabilities, model governance, and a genuine working relationship between research and operations.
The practical recommendation is therefore neither to declare a revolution nor to dismiss the approach. Choose one valuable problem, assemble a multidisciplinary team, reproduce the research results first, and only then test real enterprise data against simpler alternatives.
If KANDy passes those tests, it will be more than another prediction engine. It could become a tool that expands an enterprise's ability to understand the systems on which it depends. That is a far more strategic capability than another incremental improvement in accuracy.