Interactive Web Applications Combine Artificial Intelligence with Modular Scientific Simulation Engines

Researchers and educators studying complex dynamical systems frequently encounter significant technical barriers when selecting computational modeling platforms, as traditional scientific software environments such as MATLAB, Mathematica, AnyLogic, and R often demand specialized training curricula before users can build or modify functional models. To address these operational hurdles, computer scientists Richard Salter and Wayne M. Getz have introduced an open, browser-based simulation ecosystem built upon their proprietary NuWebKit architecture. Published in Frontiers in Systems Biology, the investigation documents an interactive computational environment that executes simulations directly within standard web browsers or on remote computing servers without requiring client software installations. The system incorporates artificial intelligence capabilities that translate plain language prompts into functional simulation routines, alongside an editable runtime framework that allows practitioners to modify mathematical formulations directly on a running interface. By decoupling the core computational engine from front-end visual components, the design aims to democratize access to advanced mathematical modeling across biological, epidemiological, and ecological disciplines.

The structural foundation of the platform relies on Runtime Alterable Module, or RAM, technology, an architectural feature developed through the authors’ earlier work on Numerus Designer and Numerus Studio. In contrast to conventional random-access memory hardware, this software mechanism allows users to dynamically insert or alter mathematical equations on the fly using a concise domain-specific syntax rather than editing underlying engine code. NuWebKit applications maintain a single, unified code base across diverse modeling domains, achieving project-specific customization through dedicated JSON configuration files termed project DNA. This centralized specification file designates interface geometry, root deployment directories, control widgets, and analytical modules. Consequently, the simulation engine can be deployed client-side in native JavaScript for routine educational models, or hosted server-side utilizing high-performance Java APIs connected via WebSockets when executing demanding mathematical workloads that exceed standard browser execution capabilities.

To guarantee computational precision across discrete and continuous modeling frameworks, the NuWebKit simulation engine implements rigorous state-management protocols during iterative clock cycles. The simulator subdivides clock pulses into distinct computational phases, beginning with a step phase that calculates prospective state values, followed by a shift phase that concurrently commits the newly computed values to the active model state, and concluding with an update phase that broadcasts data to the visual dashboard. The system organizes variable updates using specialized store and stock data structures to maintain strict concurrency across distributed model components. For continuous systems defined by differential equations, stock structures handle numerical integration through a fourth-order Runge-Kutta algorithm, designated as RK4, paired with user-specified integration time intervals. Conversely, systems governed by discrete difference equations employ Euler integration with unitary step parameters, ensuring consistent temporal progression regardless of whether underlying model logic is deterministic or stochastic.

Beyond classical compartmental calculus, the platform provides comprehensive native infrastructure for spatial and agent-based simulations by adapting foundational motifs from established platforms like NetLogo. In this context, mobile entities termed turtles navigate across spatial environments divided into discrete cells known as patches, overseen by coordinated base classes that regulate state transitions across clock cycles. Unlike rigid Cartesian grids that introduce geometric distance distortions between diagonal and rectilinear neighbors, NuWebKit incorporates hexagonal patch arrays that preserve equal spatial relationships among adjacent coordinates. Developers can configure spatial topologies to feature hard boundaries where agents bounce like billiard balls, or continuous toroidal geometries that wrap edge boundaries seamlessly. During execution cycles, targeted optimization algorithms monitor internal state modifications, restricting iterative computational updates exclusively to active cells and proximate entities, thereby preserving processing bandwidth during large-scale ecological or cellular simulations.

The user-facing environment consolidates interactive operation into a responsive dashboard complemented by an expandable slide-out drawer housing configuration controls, documentation panels, and preset managers. Through the drawer interface, users can store up to twenty-six distinct runtime operational configurations as persistent browser presets, enabling rapid navigation between diverse parameter suites without reloading the host page. Preset zero remains permanently hardcoded to default factory parameters, while subsequent memory registers retain custom parameter valuations, mathematical variations, and widget arrangements locally across distinct browser sessions. For specialized analyses, dashboard tabs extend base model functionality by launching dedicated analytical engines. These supplementary platforms include automated parameter fitting engines employing the Nelder-Mead simplex algorithm against uploaded empirical datasets, sensitivity analysis displays plotting variance across repeated parametric sweeps, and Monte Carlo stochastic iteration engines generating probability distributions and extinction timelines.

At the center of user-driven customization is NuML, a specialized domain-specific language designed by the researchers to encode mathematical models without forcing users to manage interface plumbing. NuML adopts JavaScript syntax augmented by two distinctive infix operators: the definition operator, written as @=, and the configuration operator, written as #=. The definition operator pairs state variables with functional mathematical roles, allowing modelers to assign time derivatives via deriv call expressions, advance sequences via next expressions, specify initial boundary values, or direct numerical outputs into graphical widgets via update directives. Concurrently, the configuration operator sets operational simulation properties, defining slider parameters, graph limits, and integration methods. NuML incorporates a rich library of mathematical primitives and random sampling functions, including normal, Poisson, binomial, and multinomial distribution generators, enabling users to construct intricate stochastic models directly within browser coding fields.

The authors integrated large language model artificial intelligence interfaces directly into the NuWebKit architecture to automate the translation of natural language modeling descriptions into validated NuML code. Supported commercial and open-weight models, including Gemini, ChatGPT, and DeepSeek, receive system instructions establishing NuML syntactic grammar and structural rules. When an operator inputs a natural language prompt, the selected language model produces two synchronized deliverables: an executable NuML script and a comprehensive technical documentation document rendered with Markdown formatting and MathJax mathematical typography. The AI workspace maintains isolated conversation threads across multiple artificial intelligence engines simultaneously, permitting researchers to pivot between distinct language models without losing context, while automatically purging conversation histories whenever users switch simulation presets to prevent parameter cross-contamination across unrelated research projects.

To evaluate the practical capabilities of the AI modeling framework, the researchers benchmarked the platform against classical problems across varying levels of mathematical complexity. In a continuous-time neurobiological test, the authors tasked artificial intelligence systems with recreating the Nobel Prize-winning Hodgkin-Huxley squid giant axon action potential model. When given concise natural language prompts, Gemini generated functional NuML code that constructed an operational dashboard with dynamic voltage sliders, alongside complete technical documentation of membrane conductance equations. To verify empirical accuracy, the authors consulted an independent Claude instance to compute expected steady-state firing frequencies under sustained electrical stimulation. The simulated NuWebKit dashboard yielded a steady firing rate of sixty-eight hertz with a peak-to-trough spike amplitude of one hundred five millivolts, matching theoretical predictions exactly.

The investigators further evaluated the artificial intelligence engine against complex ecological scenarios featuring demographic stochasticity and discrete harvesting mechanics. Using a structured Markdown prompt specifying age-structured populations, sex-ratio allocation, density-dependent recruitment, and binomial survival, the software generated a multi-dimensional Leslie matrix model with four distinct age classes. The resulting simulation environment incorporated functional sensitivity graphs that computed total population variance across parameter sweeps. In a companion population viability analysis prompt, DeepSeek generated a discrete-time stochastic model tracking adult female persistence subject to Ricker density-dependent recruitment, binomial survival, and Poisson-distributed harvest pressure. Dashboard runs across five hundred to one thousand stochastic iterations produced frequency histograms of population extinction times, proving that generative models could reliably reproduce complex probabilistic population dynamics from structured text specifications alone.

To support rigorous external data pipelines, NuWebKit incorporates flexible data archiving protocols and interoperability bridges that interface with standard scientific programming environments. The platform provides three distinct data export tiers: localized comma-separated value downloads from individual dashboard widgets, consolidated multi-display comma-separated value archiving, and a high-fidelity historian module that streams granular agent transition logs directly into external SQL database architectures. Furthermore, through an integrated local bridge termed RConnect, NuWebKit applications can operate as virtual packages inside the R statistical ecosystem. External R scripts transmit operational instructions across local WebSocket bridges using a remote procedure language, allowing statistical analysts to orchestrate automated batch simulations, execute Markov chain Monte Carlo parameter estimations, and retrieve simulation matrices directly into standard analytical data frames.

In evaluating the physical limitations and practical boundaries of their framework, the researchers reported that client-side performance varies according to hardware resources and model architectures. Benchmark tests conducted on a personal computer equipped with sixteen gigabytes of memory demonstrated that lightweight agent-based simulations could reliably support fifty thousand to one hundred thousand entities, though execution efficiency began degrading once populations surpassed fifteen thousand active agents, while complex landscape models successfully navigated dozens of mobile entities over grids comprising one million spatial cells. The authors emphasized that while artificial intelligence tools significantly lower the entry barrier for building simulation software, users must remain vigilant regarding model veracity by carefully scrutinizing mathematical formulations and testing model behavior across varied parameters. By uniting automated code synthesis, runtime modularity, and cross-platform statistical connectivity, NuWebKit establishes a versatile foundation for advancing dynamical systems education and computational research.

Subject of Research: Browser-based dynamical systems simulation modeling utilizing artificial intelligence and modular runtime architectures

Article Title: AI-enabled, user-modifiable, systems simulation web apps

Article References: Salter, R., & Getz, W. M. (2026). AI-enabled, user-modifiable, systems simulation web apps. Frontiers in Systems Biology, 6, Article 1964116. https://doi.org/10.3389/fsysb.2026.1964116

Image Credits: AI Generated

DOI: 10.3389/fsysb.2026.1964116

Keywords: computational biology, systems simulation, NuWebKit, artificial intelligence, dynamical systems, agent-based modeling, mathematical demography, open science software, AI-enabled, user-modifiable, systems, simulation