AI-powered optimization of the green supply chain
Logistics accounts for 8% of global CO2 emissions. A company with a revenue of $500 million spends $12–18 million annually on transportation and generates 45,000 tCO2e in Scope 3, Category 4 (upstream), and Category 9 (downstream). A route optimizer without ESG constraints already exists—it minimizes costs. The challenge: add a carbon budget as a hard constraint without increasing transportation costs by more than 3–5%.
Multi-objective route optimization
Standard VRP → Green VRP
Vehicle Routing Problem with CO2 constraint (GVRP). Add to the objective function:
# Multi-objective: minimize cost + alpha * CO2
# subject to: total_CO2 <= carbon_budget
def objective(routes, alpha=0.15):
total_cost = sum(route_cost(r) for r in routes)
total_co2 = sum(route_emissions(r) for r in routes)
return total_cost + alpha * co2_penalty(total_co2)
def route_emissions(route):
# Emission factor зависит от типа ТС, загрузки, топлива
# HBEFA 4.2 (Handbook Emission Factors for Road Transport)
emissions = 0
for leg in route:
ef = emission_factor(vehicle_type=leg.vehicle,
load_factor=leg.load/leg.capacity,
road_type=leg.road_category,
fuel_type=leg.fuel) # кг CO2/км
emissions += leg.distance_km * ef
return emissions
OR-Tools (Google) for base VRP, Bayesian Optimization for alpha tuning. On the test dataset with 380 delivery points and 45 vehicles: 18% CO2 reduction with a 2.3% cost increase vs. pure cost VRP.
Modal choice
Road vs. Rail vs. Sea: for interregional and international transportation. An ML classifier based on features (distance, deadline, weight, cargo type, transit countries) recommends the optimal mode based on CO2 emissions. Rail is 8-10 times cleaner than truck on comparable routes. Optimization of intermodal hubs: Integer Programming + historical delay statistics.
Supplier Carbon Footprint Assessment and Selection
Carbon Scoring Supplier
Each supplier receives a carbon score: proprietary PCF (Product Carbon Footprint) data + calculated data from EXIOBASE EEIO + industry benchmarks. Given comparable price and quality, the supplier with the best score is given priority.
ML component: Scope 3 Cat 1 emissions prediction for suppliers without PCF data. XGBoost regression on features: country, industry (NACE code), company size, revenue. RMSE $12/tCO2e vs. EEIO baseline $28/tCO2e with at least partial data.
Nearshoring analysis
The ML model estimates trade-offs: nearshoring 500 km to the supplier (higher cost) vs. offshore 8,000 km (lower cost, +380 tCO2e/year). The full cost includes the shadow carbon price ($75/tCO2e), supply chain risk score, and lead time.
Optimization of warehouse logistics
Slotting optimization with energy criterion
The location of goods in the warehouse affects forklift mileage and energy consumption. The Reinforcement Learning agent optimizes the slotting policy by fast-moving SKUs closer to the shipping gate. The warehouse has 25,000 SKUs: forklift mileage is reduced by 16%, energy consumption is reduced by 11%.
Load optimization
Maximizing vehicle load factor: 3D bin packing (LLM-enhanced heuristics) for optimal stowage. Increasing load factor from 71% to 84% = fewer trips = less CO2. Tools: Google OR-Tools 3D knapsack, PackPy.
Monitoring and carbon accounting
Each shipment generates an emissions record in real time: integration with TMS (SAP TM, Oracle TMS, Freight Tiger) + telematics (Samsara, Geotab API) for actual mileage vs. planned mileage. A discrepancy of >10% triggers recalculation.
Dashboard: emissions by carrier, by lane, by product category, by quarter. Drill down to a specific flight. Export to the ESG system for automatic replenishment of Scope 3 Cat 4 data.
Alternative fuels
Electric fleet planning: An ML model analyzes route history and determines which trips are suitable for electrification (based on range, charging time, and availability of charging infrastructure). For a fleet of 120 trucks, the optimal EV share is 34% with available charging infrastructure.
HVO (Hydrogenated Vegetable Oil), LNG, CNG – assessment of economics and CO2 reduction for each option, taking into account current contracts and infrastructure.
Development time: 3–7 months for Green VRP + supplier scoring. A full platform with TMS integration and real-time monitoring: 8–12 months.