ARTERY: crisis simulation software

From a plain‑language description of a geopolitical shock, ARTERY forecasts how European freight, energy and dual‑use trade reorganize — the new routes, the grey and illicit channels, the prices, and the enforcement that chases them.

The figures and scenario shown here are a demonstrator, not a live intelligence product. AAIC built the engine; the interface in these recordings is a technical demonstrator of it, not a production UI.

The brief

Intelligence analysts answer a recurring question under deadline: if this happens, how does the movement of goods across Europe change—and what does it cost? “This” is rarely tidy: a strait is closed, a corridor is sanctioned, a price cap is imposed, a border shuts. The true answer is multidimensional—routes shift, modes substitute, prices and insurance move, and a fraction of trade slips into grey re‑export or a dark shadow fleet.

ARTERY exists to make that reasoning explicit, quantitative and legible. An analyst states a situation and a set of conditions; the system simulates the reorganization of the network over time and presents it as geography, markets, a route ledger, a flow network and a compliance picture.

Fig. 1Stating the world. A free‑text situation, a list of trigger events, and a rule set (embargoes, a price cap, dual‑use export bans, border closures) are compiled into a constraint model over a network of ports, rail gauges, pipelines and road crossings.

What the system does

ARTERY models Europe and its approaches as a multi‑modal network: nodes are ports, rail break‑of‑gauge points, pipeline injection points, road crossings and ship‑to‑ship zones; arcs are corridors with a mode, a capacity, a transit time and a cost. A scenario perturbs this network—severing arcs, capping prices, raising war‑risk premiums—and the system computes how flow redistributes, period by period, as actors adapt.

It distinguishes four channel classes that matter to the analyst: legal direct trade, rerouted legal trade (e.g. crude sold below a price cap), grey transshipment (legal‑but‑evasive parallel imports relabelled through third countries), and dark flows (shadow‑fleet crude, dark‑AIS, smuggling). The same engine prices each channel and tracks the enforcement that erodes them.

Fig. 2Reorganization on real geography. Direct Baltic and pipeline corridors sever; flow re‑forms as a grey network through the Caucasus, Central Asia, Türkiye, the Gulf and the Balkans, and a dark fleet running ship‑to‑ship off Greece.

Why this is hard

Underneath the animation is a genuinely difficult optimization problem, and it is difficult in several ways at once:

Combinatorial. Choosing which corridors carry which commodity, and which new routes to open, is a fixed‑charge multi‑commodity flow problem—NP‑hard, with binary route‑opening decisions over a large arc set.
Multi‑objective. Actors do not minimize one thing. They trade off landed cost, transit time, capacity risk, legal exposure and probability of detection. There is no single optimum—only a Pareto surface.
Non‑convex and coupled. Price is not exogenous. Rerouting moves volumes, volumes move prices and insurance, prices change what is worth rerouting. The system must reach an equilibrium, not a one‑shot route.
Adversarial. Grey and dark channels form because of the constraints. Enforcement and evasion are a leader–follower game; every interdiction policy induces a new best response.
Uncertain and dynamic. Elasticities, spare capacity, enforcement reach and adversary intent are unknown and time‑varying. The answer is a distribution that evolves, not a number.

No single algorithm is adequate to all five. ARTERY is therefore a hybrid: exact mathematical programming where structure must be respected, and learned models where the world is too large, too fast or too uncertain to solve from scratch.

The mathematical core

The base layer is a fixed‑charge multi‑commodity flow over the corridor graph G = (N, A), with commodities k, continuous flows x and binary route‑opening variables y:

\begin{aligned} min & \sum_{k} \sum_{(i, j) \in A} c_{i j}^{k} x_{i j}^{k} + \sum_{(i, j)} f_{i j} y_{i j} \\ s.t. & \sum_{j} x_{i j}^{k} - \sum_{j} x_{j i}^{k} = b_{i}^{k} & conservation, ∀ i, k \\ \sum_{k} x_{i j}^{k} \leq u_{i j} \cdot y_{i j} & capacity, arc open \\ x_{i j}^{k} = 0 \forall (i, j) \in Prohibited (s) & legality under scenario s \\ x \geq 0, y \in {0, 1} \end{aligned}

This is solved by branch‑and‑cut, with column generation proposing candidate routes (paths and transshipment chains) only as they become attractive—essential because the set of conceivable grey routings is astronomically large. Over the legal arc set, the relaxation is a clean min‑cost flow; the binaries capture the discrete act of opening a channel.

Because actors weigh several incommensurable goals, the objective is really a vector, and the solution is a Pareto set rather than a point:

\begin{aligned} minimize & F (x) = (cost (x), time (x), detection‑risk (x), legal‑exposure (x)) \\ seek & P = {x : no feasible x^{'} dominates x} \end{aligned}

ARTERY traces P with the ε‑constraint method for small instances and an evolutionary multi‑objective search (NSGA‑II–style) for large ones, then lets the scenario’s posture (least‑cost, least‑time, least‑detection) pick the operating point.

Routing is not enough, because prices respond. Each period the network is driven to a spatial price equilibrium — a Wardrop‑type complementarity condition in which a corridor carries flow only when its landed cost exactly closes the price gap it spans:

for every arc (i, j) : 0 \leq x_{i j} ⊥ (c_{i j} + κ_{i j} - (μ_{j} - μ_{i})) \geq 0

μ : nodal price κ_{i j} : risk / insurance premium on the arc

Solved as a mixed complementarity problem (a variational inequality), this is what makes the demonstrator’s markets move coherently: the Urals discount, the war‑risk premium and the grey‑channel margin are not scripted independently—they are the prices that clear the rerouted network.

Fig. 3Prices as an equilibrium. As the $60 cap binds, the benchmark–discount spread, insurance, markup and leakage are read off the cleared network rather than assumed.

Where AI is required

Exact programming is correct but slow, and it cannot read intent, learn behaviour or price the unknown. ARTERY pairs it with learned components:

Language model for ingestion. The free‑text situation and conditions are parsed by an LLM into a typed constraint set—actors, instruments, prohibited arcs, parameter changes—the structured input the solver consumes. This is the “insert a world situation” surface.
Graph‑neural‑network surrogate. A GNN is trained on thousands of solved instances to approximate the optimizer, f_φ(s) ≈ argmin_x C(x;s). It returns a near‑instant answer for interactive what‑ifs; the exact solver then certifies only the handful of decisions that matter (learning‑guided branch‑and‑bound).
Multi‑agent reinforcement learning. Shippers, brokers and fleet operators are agents that learn to reroute and to stand up grey channels. Their adaptation—and the formation of shadow logistics—is the emergent output of RL policies, not a hand‑written rule.
Stackelberg enforcement model. Interdiction is a bilevel game: the enforcer commits to a detection policy θ; the followers best‑respond by minimizing landed cost inclusive of expected seizure.

\begin{aligned} {max}_{θ} & leakage (θ) \cdot value - budget (θ) & enforcer — leader \\ s.t. & x^{*} (θ) \in {argmin}_{x} [landed (x) + p (θ) \cdot seizure (x)] & shippers — followers \end{aligned}

The interplay—symbolic constraints that must hold, neural estimators for the quantities that can only be learned—is deliberately neuro‑symbolic: the AI proposes structure (which channels could exist, how demand responds); the optimizer disposes (whether they are feasible and what they cost).

Handling uncertainty

The inputs that matter most—price elasticities, spare capacity, enforcement reach, adversary willingness—are precisely the ones least known. ARTERY treats uncertainty as first‑class rather than averaging it away.

Scenario‑based stochastic programming. Decisions are optimized against a sampled ensemble of futures ξ drawn from priors over the uncertain parameters.
Robust / worst‑case bounds. For high‑consequence questions it also reports the adversarial optimum over an uncertainty set U, so an analyst sees the downside, not only the expectation.
Monte‑Carlo propagation. Thousands of network solves turn input distributions into output distributions—fan charts and confidence bands on prices, leakage and transit time.
Bayesian updating. As real intelligence arrives (an observed seizure, a re‑export figure), posteriors over elasticities and detection rates are updated, narrowing the forecast.

expected {min}_{x} E_{ξ} [cost (x, ξ)] robust {min}_{x} {max}_{ξ \in U} cost (x, ξ)

ξ = (elasticities, capacities, detection rates, intent) — uncertain, time‑varying

The result is presented as a range, not a false point estimate. The real claim ARTERY makes is not “this will happen” but “under these conditions, the network most plausibly reorganizes like this, within these bounds, and here is what would change the answer.”

Outputs the analyst works from

Every run produces the same coherent picture across views: the map (geography of reorganization), the markets board (the cleared prices), the route ledger (every corridor with distance, transit, cost and the live markup it carries), the flow network (throughput by channel and the load on each transshipment hub), and the compliance view (shadow‑fleet registry, designations and interdictions). They are not separate reports—they are projections of one underlying equilibrium.

Fig. 4From enforcement to cost. The compliance picture—tracked shadow‑fleet vessels, designations, seizure rate—and the full route ledger, where each lane’s landed cost and markup follow from its status in the cleared model.

In short

ARTERY turns a described crisis into a quantified, time‑resolved forecast of how European transport reorganizes. It does so by refusing the false choice between mathematics and AI: a fixed‑charge multi‑commodity flow with column generation for routing, multi‑objective search for the trade‑offs people actually make, a spatial‑price complementarity for the markets, and a Stackelberg game for evasion versus enforcement—wrapped in learned components for ingestion, speed, behaviour and uncertainty. It is built to inform judgement under deadline, with its uncertainty shown rather than hidden.