Exploring Super Automation in JD Supply Chain: Architecture, Applications, and Future Outlook

The article introduces the concept of supply chain super automation, which aims to achieve overall automation and optimization of all supply chain stages by leveraging machine learning, operations research, and intelligent decision making.

JD's intelligent supply chain serves three major segments: JD's self‑operated business, POP merchants on the JD platform, and SaaS solutions for key accounts and brands, with the goal of improving procurement efficiency, reducing inventory turnover, and lowering fulfillment costs.

Challenges include a massive SKU count (over 10 million), diverse product categories, six logistics networks covering 1,400+ warehouses, and multiple cooperation models (self‑operated, direct‑factory, FCS).

JD divides the supply chain into planning, procurement, inventory, marketing, and fulfillment, focusing on core technical capabilities such as product matching, intelligent diagnosis, real‑time computing, large‑scale optimization, and supply‑chain simulation.

The four key business processes—planning, procurement, inventory, marketing, and fulfillment—are illustrated with images of network topology and workflow diagrams.

Exploration and Application

1. Demand Forecasting Automation

Demand management is the foundation of supply‑chain management, involving intelligent forecasting and inventory planning. Challenges include model selection for diverse items, information asymmetry (e.g., missing future promotion plans), and rapidly diversifying consumer demand across channels.

Solution 1: Apply business constraints to handle cases like clearance promotions.

Solution 2: Use a routing model to match each SKU and sales channel with the best base model from a large model pool.

Solution 3: Incorporate reinforcement learning to embed inventory‑turnover metrics into the reward function.

2. Information Asymmetry

Marketing and planning information are often unavailable; JD builds a marketing‑indicator system to predict future promotions and uses algorithmic estimation of sales plans to mitigate missing data, ensuring robust predictions despite imperfect inputs.

3. Algorithm Optimization

Differences between B2C and B2B scenarios require distinct models; B2B demands handling smaller order volumes but larger spikes, leading to specialized deep‑learning architectures for baseline and peak demand sequences.

2. Procurement Automation

Replenishment models (e.g., TIBPA, end‑to‑end) address forecast uncertainty and business constraints such as shipping costs, supplier capacity, and shelf life. Integer programming is used for parameter recommendation, and a feedback loop evaluates replenishment outcomes.

3. Intelligent Allocation

Allocation aims to move goods closer to users, balancing inventory cost and service level. Three allocation patterns are described: R‑F (regional to local), R‑R (regional to regional), and F‑R (local to regional), each with selection and model‑calculation steps.

4. Inventory Clearing Automation

Clearing slow‑moving inventory involves four steps: identification, diagnosis, decision recommendation (e.g., promotion, F2R, R2R), and execution tracking, all driven by optimization techniques.

5. Integrated Decision‑Making

Each module (forecasting, replenishment, allocation, clearing) pursues local optimality; the integrated system seeks a global optimum by balancing inventory levels, procurement, and clearance actions.

Future Outlook

Super automation progresses through four levels: assisted decision automation, single‑process automation, cross‑process automation, and full super automation, each requiring advances in both business and technical domains.

Q&A

Q1: How is forecasting split to SKU granularity? A: By estimating future discounts from sales plans, linking historical sales and traffic, and using these inputs to predict SKU‑level sales.

Q2: How is parameter recommendation performed? A: Formulated as an integer‑programming optimization problem with business constraints (stock‑on‑hand, turnover, capacity, scheduling) to obtain optimal parameters.

Images illustrating network topology, model pipelines, and process flows are embedded throughout the article.