go-pflow

README

go-pflow

Note: go-pflow is not an AI or machine learning library. It implements structural, dynamical computation based on Petri nets and ODE simulation. See How go-pflow Differs From Modern AI for details.

A Go library for Petri net modeling, ODE simulation, process mining, and predictive monitoring.

Port of the JavaScript pflow.xyz library with additional features:

• Learn process models from event logs
• Predict case completion times in real-time
• Fit parameters using Neural ODE-ish approaches
• Continuous state machine operation with triggers

🚀 Quick Navigation

New to go-pflow? Start here:

• Installation - Get up and running in 2 minutes
• Quick Start - Your first simulation
• 📚 Documentation Hub - Comprehensive learning materials
• 🎯 Examples - Working demos and tutorials

Looking for something specific?

• Package Reference - API documentation
• CLI Tool - Command-line interface
• Process Mining - Learn from event logs
• Architecture - System design

Features

Core Modeling

• Petri Net Modeling: Define places, transitions, and arcs with fluent Builder API
• JSON Import/Export: Load and save Petri nets in JSON-LD format compatible with pflow.xyz
• ODE Simulation: Convert Petri nets to ODEs using mass-action kinetics with multiple solvers (Tsit5, RK45, implicit methods)
• Equilibrium Detection: Automatic steady-state detection with configurable tolerances

Higher-Level Abstractions

• Metamodel DSL: Declarative schema language for defining state machines as Petri nets with guard expressions
• Solidity Codegen: Generate audit-ready Solidity smart contracts from metamodel schemas
• Workflow Framework: Task dependencies, resource management, SLA tracking, and case monitoring
• State Machine Engine: Hierarchical states, parallel regions, guards, and actions compiled to Petri nets
• Actor Model: Message-passing actors with Petri net behaviors, signal bus, and middleware

Analysis & Optimization

• Reachability Analysis: State space exploration, deadlock detection, liveness analysis, P-invariants
• Sensitivity Analysis: Parameter impact ranking, gradient computation, grid search optimization
• Hypothesis Evaluation: Parallel move evaluation for game AI and decision making
• Caching: Memoization for repeated simulations with LRU eviction

Process Mining

• Process Discovery: Alpha Miner and Heuristic Miner algorithms
• Rate Learning: Learn transition rates from event log timing data
• Predictive Monitoring: Real-time case tracking with SLA prediction and alerting

Learning & Visualization

• Neural ODE-ish Learning: Fit learnable transition rates to observed data
• SVG Visualization: Petri nets, workflows, state machines, and simulation plots

Installation

go get github.com/pflow-xyz/go-pflow

Quick Start

package main

import (
    "fmt"
    "github.com/pflow-xyz/go-pflow/petri"
    "github.com/pflow-xyz/go-pflow/solver"
    "github.com/pflow-xyz/go-pflow/plotter"
)

func main() {
    // Create a simple Petri net: A -> B
    net := petri.NewPetriNet()
    net.AddPlace("A", 100.0, nil, 0, 0, nil)
    net.AddPlace("B", 0.0, nil, 0, 0, nil)
    net.AddTransition("convert", "default", 0, 0, nil)
    net.AddArc("A", "convert", 1.0, false)
    net.AddArc("convert", "B", 1.0, false)

    // Set up simulation
    initialState := net.SetState(nil)
    rates := map[string]float64{"convert": 0.1}
    prob := solver.NewProblem(net, initialState, [2]float64{0, 20}, rates)

    // Solve
    sol := solver.Solve(prob, solver.Tsit5(), solver.DefaultOptions())

    // Plot results
    svg, _ := plotter.PlotSolution(sol, nil, 800, 600,
        "A → B Conversion", "Time", "Amount")

    fmt.Println("Final state:", sol.GetFinalState())
    // Save SVG...
}

📚 Documentation

Comprehensive documentation for learning and reference:

📖 Concepts & Tutorials

Learn the fundamentals and build real systems:

• Documentation Hub - Complete learning path for undergraduates and practitioners
  • How go-pflow Differs From Modern AI - What go-pflow is (and isn't)
  • Petri Nets Explained - What they are and why we use them
  • ODE Simulation - Why differential equations for discrete systems
  • Process Mining - Discover processes from event logs
  • Predictive Monitoring - Real-time prediction and SLA detection
  • Getting Started Tutorial - Your first working system
  • Mathematics Reference - Equations, derivations, and theory

📦 Package Documentation

Detailed API and implementation documentation:

• metamodel - Declarative schema DSL for state machines
  • S-expression DSL and fluent builder API
  • Guard expressions for transition preconditions
  • Petri net execution semantics
• codegen/solidity - Solidity smart contract generation
  • Generate contracts from metamodel schemas
  • Guard expressions to require statements
  • ERC token standards (ERC-20, ERC-721, ERC-1155)
• eventlog - Event log parsing, analysis, and statistics
• mining - Process discovery and rate learning from logs
• monitoring - Real-time case tracking and prediction
• schema - JSON format specifications and validation

🛠️ Tools & CLI

• pflow CLI - Command-line tool for simulation, analysis, and plotting
  • AI-native design for seamless integration with assistants
  • Structured JSON output optimized for machine consumption
  • Full simulation, analysis, and visualization pipeline

🤖 AI Assistant Guide

• CLAUDE.md - Comprehensive guide for AI assistants (Claude, etc.)
  • Evolutionary development approach (logs → mining → validation → features)
  • When to use Petri nets for different problem types
  • Problem-specific patterns (games, optimization, constraints, epidemics)
  • Solver tuning and performance optimization
  • Code templates and idioms

📊 Project Documentation

• ROADMAP - Development status, completed features, and future plans
• ACCOMPLISHMENTS - Key achievements and milestones
• RESEARCH PAPER OUTLINE - Academic context and contributions
• PROCESS MINING DIRECTIONS - Process mining methodology and approach

🎯 Examples

Complete working demonstrations organized by complexity and purpose. See examples/README.md for detailed comparisons and teaching progression.

Getting Started Examples

examples/basic/ - Foundation concepts

• Simple workflow and producer-consumer patterns
• Sequential token flow and resource management
• Your first Petri net simulation
• Run: cd examples/basic && go run main.go

Blockchain & Smart Contracts

examples/erc/ - ERC Token Standards ⭐

• Define token standards as Petri net schemas
• ERC-20 (fungible), ERC-721 (NFT), ERC-1155 (multi-token)
• Generate Solidity contracts from schemas
• Guard expressions, flows, and invariants
• Run: go run ./examples/erc

Integration Example (Kitchen Sink)

examples/coffeeshop/ - Comprehensive demo ⭐

• Actor pattern for high-level orchestration
• Petri nets for inventory management
• Workflows for order processing with SLAs
• State machines for equipment/staff/customer states
• ODE simulation for capacity planning
• Process mining for analyzing event logs
• Run: cd examples/coffeeshop/cmd && go run main.go

Process Mining & Monitoring

examples/eventlog_demo/ - Event log analysis

• Parse CSV event logs
• Extract timing statistics
• Analyze process behavior
• Run: cd examples/eventlog_demo && go run main.go

examples/mining_demo/ - Process discovery

• Discover process models from logs
• Learn transition rates automatically
• Compare discovered vs actual behavior
• Run: cd examples/mining_demo && go run main.go

examples/monitoring_demo/ - Real-time prediction ⭐

• Live case tracking and prediction
• SLA violation detection with advance warning
• Complete hospital patient flow example
• Run: cd examples/monitoring_demo && go run main.go

examples/incident_simulator/ - IT incident management

• Simulate incident lifecycle (detection → resolution)
• Predict completion times and SLA violations
• Real-world IT operations example
• See REGRESSION_TEST_EXPLAINED.md

Machine Learning Examples

examples/neural/ - Neural ODE-ish learning

• Fit learnable rates to observed data
• A → B decay with rate recovery
• SIR model with multiple learnable parameters
• Compare true vs fitted trajectories
• Run: cd examples/neural && go run main.go

examples/dataset_comparison/ - Model calibration

• Fit models to real-world datasets
• Compare different parameterizations
• Validate model accuracy

Game AI Examples

examples/tictactoe/ - Perfect play game AI

• Minimax algorithm with ODE-based evaluation
• Game tree exploration
• Pattern recognition for win detection
• Compare strategies: random, pattern, minimax, ODE
• Complexity: 5,478 legal positions, solved game
• Run: cd examples/tictactoe && go run ./cmd

examples/nim/ - Optimal game strategy

• Discrete state space modeling
• Optimal strategy based on Grundy numbers
• ODE-based position evaluation
• Complexity: Linear chain, provably optimal
• Run: cd examples/nim && go run ./cmd

examples/connect4/ - Complex pattern recognition

• 69 window patterns per board state
• Threat detection and blocking
• Lookahead search (minimax-lite)
• Complexity: ~10^13 legal positions, 130 places, 222 transitions
• Petri Net Model: Full board state + win detection in net structure
• See MODEL_EVOLUTION.md
• Run: cd examples/connect4 && go run ./cmd

Puzzle & Constraint Satisfaction Examples

examples/sudoku/ - Constraint satisfaction modeling

• Sudoku rules as Petri net structure
• Constraint propagation via transitions
• ODE analysis for solution detection
• Colored Petri nets for digit representation
• See ODE_ANALYSIS.md
• Run: cd examples/sudoku/cmd && go run *.go

examples/chess/ - Classic chess problems ⭐

• N-Queens: Place queens without attacks (backtracking + ODE)
• Knight's Tour: Hamiltonian path with Warnsdorff heuristic
• N-Rooks: Permutation matrix placement
• ODE-based move evaluation with optimized solver parameters
• See examples/chess/README.md
• Run: cd examples/chess/cmd && go run *.go --problem=queens

Optimization Examples

examples/knapsack/ - Combinatorial optimization

• 0/1 knapsack problem as Petri net with mass-action kinetics
• Transition rates encode value/weight efficiency preferences
• Exclusion analysis: disable items to measure sensitivity
• Continuous relaxation of discrete optimization
• Same pattern as game move evaluation (disable → observe)
• Run: cd examples/knapsack/cmd && go run *.go

Visualization Examples

examples/visualization_demo/ - SVG generation demo

• Petri net visualizations (SIR model, producer-consumer)
• Workflow diagrams (approval, parallel, incident management)
• State machine diagrams (traffic light, order status, media player)
• Run: make run-visualization

Documentation & Analysis

• examples/VISUALIZATIONS.md - Gallery of visualization examples
• examples/FEATURE_REVIEW.md - Feature comparison across examples
• examples/PARITY_STATUS.md - Implementation status matrix

Example Comparison

Example	Type	Complexity	Key Concepts	Best For Learning
basic	Workflow	Simple	Token flow, sequential processes	Petri net fundamentals
erc	Blockchain	Medium	Token standards, guards, Solidity codegen	Smart contract modeling ⭐
coffeeshop	Integration	Complex	Actors, workflows, state machines, mining	All features together ⭐
eventlog_demo	Analysis	Simple	CSV parsing, statistics	Event log basics
mining_demo	Discovery	Medium	Process discovery, rate learning	Process mining
monitoring_demo	Real-time	Medium	Prediction, SLA detection	Production systems ⭐
incident_simulator	Operations	Medium	IT workflows, SLA prediction	Real-world processes
neural	ML	Medium	Parameter fitting, learning	Data-driven modeling
dataset_comparison	Calibration	Medium	Model fitting, validation	Model selection
tictactoe	Game AI	Medium	Minimax, perfect play	Game theory
nim	Game Theory	Medium	Optimal strategy, discrete states	Mathematical modeling
connect4	Game AI	Complex	Pattern recognition, lookahead	Advanced AI techniques
sudoku	Puzzle	Medium	Constraint satisfaction, colored nets	CSP modeling
chess	Puzzle/AI	Complex	N-Queens, Knight's Tour, ODE heuristics	Classic algorithms ⭐
knapsack	Optimization	Medium	Mass-action kinetics, exclusion analysis	Combinatorial optimization
visualization_demo	Visualization	Simple	SVG rendering, workflows, state machines	Model documentation

Package Structure

Comprehensive API reference for each package:

Core Packages

petri - Core Petri Net Data Structures

Defines the fundamental building blocks with both explicit and fluent APIs:

// Explicit construction
net := petri.NewPetriNet()
net.AddPlace("p1", initialTokens, capacity, x, y, labelText)
net.AddTransition("t1", role, x, y, labelText)
net.AddArc(source, target, weight, inhibitTransition)

// Fluent Builder API
net := petri.Build().
    Place("S", 999).Place("I", 1).Place("R", 0).
    Transition("infect").Transition("recover").
    Arc("S", "infect", 1).Arc("I", "infect", 1).Arc("infect", "I", 2).
    Arc("I", "recover", 1).Arc("recover", "R", 1).
    Done()

// With rates
net, rates := petri.Build().
    SIR(999, 1, 0).
    WithCustomRates(map[string]float64{"infect": 0.3, "recover": 0.1})

Key types:

• Place - State locations that hold tokens
• Transition - Events that consume/produce tokens
• Arc - Directed connections between places and transitions
• PetriNet - Complete net structure
• Builder - Fluent API for net construction

parser - JSON Import/Export

Load and save Petri nets in pflow.xyz JSON format:

// Import
net, err := parser.FromJSON(jsonData)

// Export
jsonData, err := parser.ToJSON(net)

solver - ODE Simulation

Converts Petri nets to ODEs using mass-action kinetics and solves them:

// Create problem
prob := solver.NewProblem(net, initialState, [2]float64{t0, tf}, rates)

// Solve with Tsit5 (5th order adaptive Runge-Kutta)
sol := solver.Solve(prob, solver.Tsit5(), solver.DefaultOptions())

// Access results
finalState := sol.GetFinalState()
timeSeries := sol.GetVariable("place_name")

Solver options:

• Dt - Initial time step (default: 0.01)
• Dtmin / Dtmax - Step size bounds
• Abstol / Reltol - Error tolerances
• Adaptive - Enable adaptive stepping (default: true)

plotter - SVG Visualization

Generate publication-ready SVG plots:

// Create plotter
p := plotter.NewSVGPlotter(800, 600)
p.SetTitle("My Plot").SetXLabel("Time").SetYLabel("Concentration")
p.AddSeries(xData, yData, "Series 1", "#ff0000")

// Render
svg := p.Render()

// Or use convenience function
svg, plotData := plotter.PlotSolution(sol, variables, 800, 600,
    title, xlabel, ylabel)

engine - State Machine Harness

For continuous simulation with condition-based triggers:

// Create engine
engine := engine.NewEngine(net, initialState, rates)

// Add rules
engine.AddRule("alert_high",
    engine.ThresholdExceeded("place_name", 100.0),
    func(state map[string]float64) error {
        fmt.Println("Threshold exceeded!")
        return nil
    },
)

// Run continuous simulation
ctx := context.Background()
engine.Run(ctx, 100*time.Millisecond, 0.1) // interval, dt
defer engine.Stop()

// Or run batch simulation
sol := engine.Simulate(duration, opts)

Condition helpers:

• ThresholdExceeded(place, value) - Trigger when value exceeds threshold
• ThresholdBelow(place, value) - Trigger when value falls below
• AllOf(...conditions) - Combine with AND
• AnyOf(...conditions) - Combine with OR

Process Mining Packages

eventlog - Event Log Analysis

Parse, analyze, and extract insights from event logs:

// Load event log from CSV
log, err := eventlog.LoadFromCSV("data.csv")

// Get statistics
stats := log.Statistics()
fmt.Printf("Cases: %d, Events: %d\n", stats.NumCases, stats.NumEvents)

// Extract timing information
timings := eventlog.ExtractTimings(log)

Full documentation: eventlog/README.md

mining - Process Discovery & Learning

Discover process models and learn parameters from event logs:

// Discover process model
net, err := mining.DiscoverProcess(log, mining.CommonPathMethod)

// Learn transition rates from timing data
rates := mining.LearnRates(net, timings)

// Create runnable simulation
prob := solver.NewProblem(net, initialState, timespan, rates)
sol := solver.Solve(prob, solver.Tsit5(), solver.DefaultOptions())

Discovery methods:

• CommonPathMethod - Find most common execution path
• SequentialMethod - Sequential activity ordering
• GraphBasedMethod - Graph-based discovery (planned)

Full documentation: mining/README.md

monitoring - Real-Time Prediction

Track live cases and predict completion times:

// Create monitor
monitor := monitoring.NewMonitor(net, rates)

// Add SLA rules
monitor.AddSLARule("4-hour-rule", 4*time.Hour, 
    func(caseID string, prediction monitoring.Prediction) {
        fmt.Printf("Case %s predicted to violate SLA\n", caseID)
    })

// Process incoming events
monitor.ProcessEvent(event)

// Get predictions for active cases
predictions := monitor.GetPredictions()

Full documentation: monitoring/README.md

Advanced Packages

learn - Neural ODE-ish Parameter Learning

Fit learnable transition rates to observed data while keeping the Petri net structure as a prior:

// Create learnable rate functions
rf := learn.NewLinearRateFunc([]string{}, []float64{0.05}, false, false)
learnProb := learn.NewLearnableProblem(
    net, initialState, [2]float64{0, 30},
    map[string]learn.RateFunc{"convert": rf},
)

// Prepare observed data
times := learn.GenerateUniformTimes(0, 30, 16)
data, _ := learn.NewDataset(times, map[string][]float64{
    "A": observedA,
    "B": observedB,
})

// Fit parameters to minimize loss
opts := learn.DefaultFitOptions()
opts.Method = "nelder-mead"
result, _ := learn.Fit(learnProb, data, learn.MSELoss, opts)

fmt.Printf("Fitted rate: %.4f\n", result.Params[0])
fmt.Printf("Final loss: %.4f\n", result.FinalLoss)

Key types:

• RateFunc - Interface for learnable rate functions
• LinearRateFunc - Linear model: k = θ₀ + Σᵢ θᵢ * state[placeᵢ]
• MLPRateFunc - Small MLP with one hidden layer
• LearnableProblem - ODE problem with learnable rates
• Dataset - Observed trajectories for training
• LossFunc - Loss function (MSE, RMSE, etc.)

Optimization methods:

• nelder-mead - Nelder-Mead simplex algorithm (default)
• coordinate-descent - Simple coordinate descent

See examples/neural/ for complete examples including SIR model parameter recovery.

stateutil - State Manipulation Utilities

Helper functions for working with state maps:

import "github.com/pflow-xyz/go-pflow/stateutil"

// Copy and modify state
newState := stateutil.Copy(state)
newState := stateutil.Apply(state, map[string]float64{"A": 10, "B": 0})

// Analyze state
total := stateutil.Sum(state)
active := stateutil.NonZero(state)
changes := stateutil.Diff(before, after)

// Compare states
if stateutil.Equal(s1, s2) { /* identical */ }
if stateutil.EqualTol(s1, s2, 1e-6) { /* within tolerance */ }

hypothesis - Move Evaluation for Game AI

Evaluate hypothetical moves and find optimal decisions:

import "github.com/pflow-xyz/go-pflow/hypothesis"

// Create evaluator with scoring function
eval := hypothesis.NewEvaluator(net, rates, func(final map[string]float64) float64 {
    return final["wins"] - final["losses"]
}).WithOptions(solver.FastOptions())

// Find best move from candidates
moves := []map[string]float64{
    {"pos0": 0, "_X0": 1},
    {"pos1": 0, "_X1": 1},
}
bestIdx, bestScore := eval.FindBestParallel(currentState, moves)

// Sensitivity analysis
impact := eval.SensitivityImpact(state)  // Which transitions matter most?

sensitivity - Parameter Sensitivity Analysis

Analyze how parameters affect outcomes:

import "github.com/pflow-xyz/go-pflow/sensitivity"

// Create analyzer
scorer := sensitivity.DiffScorer("wins", "losses")
analyzer := sensitivity.NewAnalyzer(net, state, rates, scorer)

// Rank parameters by impact
result := analyzer.AnalyzeRatesParallel()
for _, r := range result.Ranking {
    fmt.Printf("%s: %+.2f impact\n", r.Name, r.Impact)
}

// Grid search optimization
grid := sensitivity.NewGridSearch(analyzer).
    AddParameterRange("infect", 0.1, 0.5, 5).
    AddParameterRange("recover", 0.05, 0.2, 5)
best := grid.Run()

cache - Simulation Caching

Memoize repeated simulations for performance:

import "github.com/pflow-xyz/go-pflow/cache"

// State cache for full solutions
stateCache := cache.NewStateCache(1000)
sol := stateCache.GetOrCompute(state, func() *solver.Solution {
    return solver.Solve(prob, solver.Tsit5(), opts)
})

// Score cache for game AI (lighter weight)
scoreCache := cache.NewScoreCache(10000)
score := scoreCache.GetOrCompute(state, computeScore)

// Check hit rate
stats := stateCache.Stats()
fmt.Printf("Hit rate: %.1f%%\n", stats.HitRate*100)

reachability - State Space Analysis

Analyze discrete state space properties:

import "github.com/pflow-xyz/go-pflow/reachability"

// Create analyzer
analyzer := reachability.NewAnalyzer(net).
    WithMaxStates(10000)

// Full analysis
result := analyzer.Analyze()
fmt.Printf("States: %d, Bounded: %v, Live: %v\n",
    result.StateCount, result.Bounded, result.Live)
fmt.Printf("Deadlocks: %d, Dead transitions: %v\n",
    len(result.Deadlocks), result.DeadTrans)

// Path finding
if analyzer.IsReachable(target) {
    path := analyzer.PathTo(target)
}

// Invariant analysis
invAnalyzer := reachability.NewInvariantAnalyzer(net)
if invAnalyzer.CheckConservation(initial) {
    fmt.Println("Net conserves total tokens")
}

Higher-Level Abstractions

workflow - Workflow Management Framework

Build and execute workflows with task dependencies:

import "github.com/pflow-xyz/go-pflow/workflow"

// Build workflow with fluent API
wf := workflow.New("approval").
    Name("Document Approval").
    Task("submit").Name("Submit").Manual().Duration(5*time.Minute).Done().
    Task("review").Name("Review").Manual().Duration(30*time.Minute).Done().
    Task("decide").Name("Approve?").Decision().Done().
    Task("approved").Name("Approved").Automatic().Done().
    Task("rejected").Name("Rejected").Automatic().Done().
    Connect("submit", "review").
    Connect("review", "decide").
    Connect("decide", "approved").
    Connect("decide", "rejected").
    Start("submit").
    End("approved", "rejected").
    Build()

// Execute workflow
engine := workflow.NewEngine(wf)
caseID := engine.StartCase(nil)
engine.CompleteTask(caseID, "submit", nil)

statemachine - Hierarchical State Machines

Build state machines that compile to Petri nets:

import "github.com/pflow-xyz/go-pflow/statemachine"

// Build state machine
chart := statemachine.NewChart("traffic_light").
    Region("light").
        State("red").Initial().
        State("yellow").
        State("green").
    EndRegion().
    When("timer").In("light:red").GoTo("light:green").
    When("timer").In("light:green").GoTo("light:yellow").
    When("timer").In("light:yellow").GoTo("light:red").
    Build()

// Create machine and process events
machine := statemachine.NewMachine(chart)
machine.SendEvent("timer")  // red -> green
machine.SendEvent("timer")  // green -> yellow

// Get underlying Petri net
net := chart.ToPetriNet()

actor - Actor Model with Message Passing

Build actor systems with Petri net behaviors:

import "github.com/pflow-xyz/go-pflow/actor"

// Build actor system
system := actor.NewSystem("my-system").
    DefaultBus().
    Actor("processor").
        Name("Data Processor").
        State("count", 0).
        Handle("data.in", func(ctx *actor.ActorContext, s *actor.Signal) error {
            count := ctx.GetInt("count", 0)
            ctx.Set("count", count+1)
            ctx.Emit("data.out", map[string]any{"processed": true})
            return nil
        }).
        Done().
    Start()

// Publish signals
system.Bus().Publish(&actor.Signal{
    Type:    "data.in",
    Payload: map[string]any{"value": 42},
})

visualization - SVG Rendering

Generate SVG visualizations for models:

import "github.com/pflow-xyz/go-pflow/visualization"

// Render Petri net
err := visualization.SaveSVG(net, "model.svg")

// Render workflow
err := visualization.SaveWorkflowSVG(workflow, "workflow.svg", nil)

// Render state machine
err := visualization.SaveStateMachineSVG(chart, "statemachine.svg", nil)

Testing

Comprehensive test coverage across all packages:

# Run all tests
go test ./...

# Run with coverage
go test ./... -cover

# Verbose output
go test ./... -v

# Test specific package
go test ./learn/... -v

Example: Test the learn package with verbose output:

go test ./learn/... -v
# All 16 tests pass with 72.7% coverage

Process Mining

go-pflow includes a complete process mining pipeline:

Event Logs → Process Discovery → Rate Learning → Simulation → Monitoring

Complete Workflow

// 1. Load event log
log, _ := eventlog.LoadFromCSV("patient_flow.csv")

// 2. Discover process model
net, _ := mining.DiscoverProcess(log, mining.CommonPathMethod)

// 3. Learn transition rates
timings := eventlog.ExtractTimings(log)
rates := mining.LearnRates(net, timings)

// 4. Create simulation
prob := solver.NewProblem(net, initialState, timespan, rates)
sol := solver.Solve(prob, solver.Tsit5(), solver.DefaultOptions())

// 5. Set up real-time monitoring
monitor := monitoring.NewMonitor(net, rates)
monitor.AddSLARule("4-hour", 4*time.Hour, alertHandler)

Key Capabilities

• Process Discovery: Automatically discover process models from event logs
• Timing Analysis: Extract activity durations and inter-arrival times
• Rate Learning: Learn transition rates from historical data
• Prediction: Forecast case completion times in real-time
• SLA Monitoring: Detect violations with 6+ hours advance warning

See examples/monitoring_demo/ for complete hospital patient flow example.

Architecture

The library is designed with modularity and layered abstractions:

petri/             Core data structures (places, transitions, arcs)
├── parser/        JSON serialization (compatible with pflow.xyz)
├── solver/        ODE solvers (Tsit5, RK45, implicit methods)
├── stateutil/     State manipulation utilities
├── learn/         Neural ODE-ish learnable rates
├── plotter/       SVG plot generation
├── engine/        Continuous simulation with triggers
│
Higher-Level Abstractions
├── metamodel/     Declarative schema DSL for state machines
├── codegen/       Code generation from schemas
│   └── solidity/  Solidity smart contract generation
├── workflow/      Task dependencies, resources, SLA tracking
├── statemachine/  Hierarchical states compiled to Petri nets
├── actor/         Message-passing actors with Petri net behaviors
│
Analysis & Optimization
├── reachability/  State space analysis, deadlock detection
├── sensitivity/   Parameter impact analysis, grid search
├── hypothesis/    Move evaluation for game AI
├── cache/         Memoization for repeated simulations
│
Process Mining
├── eventlog/      Event log parsing and analysis
├── mining/        Alpha Miner, Heuristic Miner, rate learning
├── monitoring/    Real-time case tracking and prediction
│
Visualization
├── visualization/ SVG rendering for nets, workflows, state machines
└── plotter/       Time series plots

cmd/pflow/         Command-line interface
examples/          Working demonstrations
docs/              Learning materials

Data Flow:

Historical Data          Real-Time Data
      ↓                       ↓
  Event Logs    →    Process Discovery    ←    Domain Knowledge
      ↓                       ↓
  Timing Analysis  →    Petri Net Model
      ↓                       ↓
  Rate Learning    →    ODE Simulation    →    Prediction
                            ↓
                     Monitoring Engine    →    Alerts & Actions

Mass-Action Kinetics

The solver converts Petri nets to ODEs using mass-action kinetics:

For a transition T with:

• Input places: P1, P2, ... with arc weights w1, w2, ...
• Output places: Q1, Q2, ... with arc weights v1, v2, ...
• Rate constant: k

The flux is: flux = k * [P1] * [P2] * ...

And derivatives:

• d[Pi]/dt -= flux * wi (consume from inputs)
• d[Qj]/dt += flux * vj (produce to outputs)

Neural ODE-ish Approach

The learn package extends the traditional ODE simulation with data-driven parameter learning:

Key Idea: The Petri net defines the structural prior (topology, mass conservation, reaction stoichiometry), while the transition rates become learnable functions k_θ(state, t) that can be fitted to observed data.

Benefits:

• Physical constraints preserved: Mass conservation, non-negativity, reaction structure
• Data-driven rates: Learn complex rate laws from observations
• Same reliable solver: Uses the existing Tsit5 adaptive integrator
• No external dependencies: Gradient-free optimization (Nelder-Mead, coordinate descent)

Use cases:

• System identification: Recover unknown rate constants from experimental data
• Model calibration: Fit parameters to match real-world trajectories
• Hybrid modeling: Combine known structure with learned components
• Adaptive control: Learn dynamics online for feedback systems

Design:

• RateFunc interface allows custom rate laws (linear, MLP, or user-defined)
• LearnableProblem wraps solver.Problem with parameterized rates
• Fit() optimizes parameters to minimize loss on observed data
• Fully compatible with existing solver and petri APIs

This approach bridges mechanistic modeling (Petri nets) with machine learning, enabling interpretable models that respect physical laws while learning from data.

Future Development

Active development focuses on:

• Engine package: Real-time state monitoring, event-driven actions, feedback control systems
• Learn package: Adjoint-based gradients, online learning, uncertainty quantification
• Monitoring package: Multi-case optimization, adaptive interventions, what-if scenarios

See ROADMAP.md for detailed development plans, priorities, and completed features.

Real-World Applications

This technology applies to any process with:

• Multiple steps - Registration, triage, processing, completion
• Timing constraints - SLAs, deadlines, service levels
• Historical data - Event logs with timestamps
• Need for prediction - Will this be late? When will it complete?

Application Domains:

• Healthcare: Patient flow, surgery scheduling, bed management, ER throughput
• Manufacturing: Production lines, quality control, delivery prediction
• Logistics: Order fulfillment, shipping, warehouse operations
• Finance: Loan processing, fraud investigation, compliance workflows
• IT Operations: Incident management, service requests, deployment pipelines
• Government: Permit processing, case management, citizen services

See examples/monitoring_demo/ for hospital patient flow and examples/incident_simulator/ for IT incident management.

Compatibility

• Go 1.23.6 or later
• JSON format compatible with pflow.xyz
• Portable - no external dependencies

License

MIT License - see LICENSE for details.

Based on the public domain implementation from pflow.xyz.

Contributing

Contributions welcome! Areas for improvement:

• Additional ODE solvers (Euler, RK4, etc.)
• Performance optimizations for large nets
• Additional visualization options
• More example models and use cases
• Enhanced engine capabilities
• Process mining algorithms (Alpha, Heuristics, Inductive)
• Additional monitoring and prediction features

See ACCOMPLISHMENTS.md for recent achievements and ROADMAP.md for planned features.

Related Resources

• pflow.xyz - Original JavaScript implementation and online editor
• RESEARCH_PAPER_OUTLINE.md - Academic context and research contributions
• PROCESS_MINING_DIRECTIONS.md - Process mining methodology
• Tsitouras, Ch. "Runge-Kutta pairs of order 5(4)..." Computers & Mathematics with Applications, 62 (2011) 770-775