- Get link
- X
- Other Apps
The purpose of this article is to provide a practical, expert-level blueprint for formulating and solving linear programming blending problems with rigorous mass balance and property specifications, so you can implement reliable blend optimization in fuels, chemicals, food, and materials.
1. Why linear programming is the default tool for blending optimization
Blending is the decision process of choosing component quantities to meet a target production amount and quality specifications at minimum cost or maximum profit.
Linear programming is widely used because many blending rules are linear in component amounts, especially mass conservation and weighted-average properties.
When your constraints remain linear, LP provides fast, globally optimal solutions and scales well to hundreds or thousands of components and specifications.
Note : If any property behaves nonlinearly under mixing, the model is not a pure LP unless you apply a valid linear approximation or a reformulation such as piecewise linearization or MILP.
2. Core LP formulation for blending
2.1 Decision variables and basic mass balance
Let i index available components, and let x_i be the mass (or volume) of component i used in the blend.
Let Q be the required total blend quantity in the same unit as x_i.
The fundamental mass balance is the single equation that conserves total blend quantity.
Sum over i of x_i = Q.If you have multiple products, you typically use x_{i,p} for component i going to product p, and enforce both component availability and product mass balances.
2.2 Objective functions that stay linear
The most common objective is minimizing raw material cost.
Let c_i be the cost per unit of component i.
Minimize Sum over i of (c_i * x_i).Profit maximization is equally linear if revenues are proportional to produced quantities and qualities are handled as constraints rather than nonlinear pricing curves.
Penalty terms for specification violations can be kept linear using slack variables if you want a “soft” specification approach.
2.3 Property constraints as weighted averages
Many blend properties can be modeled as linear constraints using weighted averages.
Let a_i be the value of a property for component i, and let A_min and A_max be required bounds for the blend property.
The blend property computed as a mass-weighted average is (Sum a_i x_i) / (Sum x_i), which becomes linear when you multiply by the known total Q.
A_min * Q <= Sum over i of (a_i * x_i) <= A_max * Q.This pattern is the backbone of LP blending for specifications such as concentration, impurity limits, nutrient content, density targets when treated as linear, and many additive constraints.
2.4 Bounds, availability, and operational limits
Real blends always include practical bounds on component usage due to inventory, contracts, process limits, or compatibility rules.
Let U_i be the maximum available quantity of component i, and L_i be a minimum usage if required.
L_i <= x_i <= U_i for each i.Share constraints are also common and linear when expressed against Q.
min_frac_i * Q <= x_i <= max_frac_i * Q.3. A complete reference template you can reuse
The LP blending model below captures the most common structure with mass balance, property bounds, and availability limits.
Decision variables: x_i >= 0 for i in components.
Given data:
Q = required blend quantity.
c_i = cost per unit.
a_{k,i} = value of property k in component i.
A_k_min, A_k_max = lower and upper limits for property k.
L_i, U_i = component bounds.
Objective:
Minimize Sum_i (c_i * x_i).
Constraints:
(1) Mass balance:
Sum_i x_i = Q.
(2) Property constraints for each property k:
A_k_min * Q <= Sum_i (a_{k,i} * x_i) <= A_k_max * Q.
(3) Component bounds:
L_i <= x_i <= U_i for all i.4. Worked example with numbers and a solver-ready matrix view
4.1 Scenario definition
Assume you need to produce Q = 1000 kg of a blend from three components, and you must meet two quality specifications while minimizing cost.
Component 1 is low cost but low quality, component 2 is moderate, and component 3 is expensive but high quality.
Property 1 can represent “active content” where higher is better and a minimum is required.
Property 2 can represent an impurity where lower is better and a maximum is required.
| Component i | Cost c_i ($/kg) | Property 1 a1_i (wt%) | Property 2 a2_i (wt%) | Availability U_i (kg) |
|---|---|---|---|---|
| 1 | 0.80 | 70 | 1.20 | 1000 |
| 2 | 1.10 | 85 | 0.60 | 700 |
| 3 | 1.60 | 98 | 0.15 | 500 |
Specifications for the final blend are Property 1 must be at least 88 wt% and Property 2 must be at most 0.50 wt%.
All x_i are in kg and must be nonnegative.
4.2 LP constraints in numeric form
Mass balance is x1 + x2 + x3 = 1000.
Property 1 minimum is 88 * 1000 <= 70 x1 + 85 x2 + 98 x3.
Property 2 maximum is 1.20 x1 + 0.60 x2 + 0.15 x3 <= 0.50 * 1000.
Availability constraints are x1 <= 1000, x2 <= 700, x3 <= 500, and x1, x2, x3 >= 0.
Objective is minimize 0.80 x1 + 1.10 x2 + 1.60 x3.
4.3 Matrix view for implementation
Many tools expect the form minimize c^T x subject to A x <= b and Aeq x = beq with x >= 0.
Using x = [x1, x2, x3]^T, you can write the constraints as follows.
Equality: [1 1 1] * x = 1000.
Inequalities:
(Property 1 minimum) -70 -85 -98 * x <= -88000.
(Property 2 maximum) 1.20 0.60 0.15 * x <= 500.
(Availability) 1 0 0 * x <= 1000.
0 1 0 * x <= 700.
0 0 1 * x <= 500.Note : Converting a “greater than or equal” constraint into solver standard form is typically done by multiplying both sides by -1, which preserves linearity and avoids mistakes.
5. Property constraints that look linear but can hide modeling traps
5.1 Mass basis versus volume basis
Weighted-average constraints are only valid when the property is additive on the same basis you are blending on.
If you blend by volume but treat properties defined on a mass basis, you can introduce systematic errors.
A practical approach is to standardize the decision variables to a single basis and convert component properties accordingly.
Note : Always document the basis of each property as “mass-based additive,” “volume-based additive,” or “non-additive,” and only apply the LP weighted-average rule to the additive cases.
5.2 Ratio specifications
Some specifications appear as ratios, such as “impurity per active,” or “component A must be at least twice component B.”
Many ratio constraints become linear when the denominator is fixed or when you cross-multiply with a known total Q.
If the denominator is variable, cross-multiplication can create bilinear terms, so you must reformulate carefully.
5.3 Non-additive properties and what to do about them
Examples of non-additive blending behavior include certain octane blending effects, vapor pressure, viscosity mixing rules, and phase behavior constraints.
Common ways to stay in a linear framework are to use conservative linear bounds, use piecewise linear approximations, or move to MILP with discrete breakpoints.
For piecewise linearization, you approximate a nonlinear property curve by segments and introduce additional variables that convexly combine segment endpoints.
This maintains tractability and gives you control over approximation error.
6. Advanced constraints used in real plants
6.1 Multiple products and component allocation
When producing multiple blends simultaneously, introduce x_{i,p} variables and add component availability constraints across all products.
For each component i: Sum over products p of x_{i,p} <= U_i.
For each product p:
Sum_i x_{i,p} = Q_p.Each product then receives its own set of property constraints.
6.2 Inventory carryover and multi-period planning
For weekly or daily planning, add time index t and include inventory variables I_{i,t}.
This produces a linear network flow structure coupled with quality constraints, which LP solvers handle efficiently.
I_{i,t+1} = I_{i,t} + supply_{i,t} - Sum_p x_{i,p,t}. 0 <= I_{i,t} <= storage_cap_{i}.6.3 Logical rules and incompatibilities
Rules like “use at most one of these two components,” or “if component A is used then component B must be used,” are not linear in pure LP form.
These are handled with binary variables and become mixed-integer linear programming.
If you can replace a logical rule with a continuous operational bound based on engineering reality, you may keep the model as an LP and gain speed and robustness.
7. How to diagnose infeasibility and improve model reliability
7.1 Common causes of infeasible blends
Infeasibility usually comes from contradictory specifications, incorrect property data, unit inconsistencies, or overly tight component bounds.
It can also come from assuming linear additivity for a property that is not additive in the relevant range.
7.2 Practical debugging workflow
Start by solving a relaxed model that includes only mass balance and bounds, then add properties one by one.
If infeasibility appears after adding one property, that constraint is the primary suspect.
Introduce slack variables to convert hard constraints into soft constraints with penalties, which helps identify which specifications drive violation pressure.
For an upper bound property: Sum_i (a_i * x_i) <= A_max * Q + s. s >= 0. Add a penalty term M * s to the objective.Note : Use large penalties thoughtfully because extremely large coefficients can create numerical conditioning problems, so scale data to similar magnitudes where possible.
7.3 Scaling and numerical conditioning
Blending models often mix costs in dollars, flow in tons, and properties in ppm or wt%, which can span several orders of magnitude.
Rescale properties to consistent units such as mass fraction, and express ppm as fraction where appropriate.
Good scaling improves solver speed and reduces the risk of misleading infeasibility due to tolerances.
8. Implementation patterns in spreadsheets and Python
8.1 Spreadsheet implementation checklist
Spreadsheets are suitable for small to medium blending problems when you set formulas carefully and use a reliable LP solver interface.
Place decision variables x_i in a column, compute Sum x_i, compute each Sum(a_{k,i} x_i), and connect those cells to constraints.
Keep a separate block for data and enforce consistent units with explicit labels.
8.2 Python implementation skeleton with a linear solver
The code below shows the structure of an LP blending model in Python-like pseudocode using standard LP concepts.
This example focuses on clarity of formulation rather than a specific library.
# Data. Q = 1000.0 components = ["C1", "C2", "C3"] c = {"C1": 0.80, "C2": 1.10, "C3": 1.60} a1 = {"C1": 70.0, "C2": 85.0, "C3": 98.0} a2 = {"C1": 1.20, "C2": 0.60, "C3": 0.15} U = {"C1": 1000.0, "C2": 700.0, "C3": 500.0}
A1_min = 88.0
A2_max = 0.50
Variables.
x[i] >= 0
Objective.
minimize sum(c[i] * x[i] for i in components)
Constraints.
sum(x[i] for i in components) == Q
sum(a1[i] * x[i] for i in components) >= A1_min * Q
sum(a2[i] * x[i] for i in components) <= A2_max * Q
x[i] <= U[i] for i in componentsNote : Treat wt% consistently as either “percent units” across all terms or convert to fraction everywhere, and do not mix the two within the same constraint.
9. Industry-specific examples of linear properties
9.1 Fuels and petroleum blending
Many fuel specs can be approximated as linear constraints, such as sulfur content, benzene limits, oxygenate percentage, and certain distillation cut constraints when represented as linear proxies.
Some key properties are non-additive and require special handling, so document which ones are treated as linear and why.
9.2 Chemical solutions and formulations
Concentration of solutes, inhibitor dosing, trace impurity caps, and additive package targets are often linear on a mass basis.
Regulatory constraints expressed as maximum mass fraction or maximum ppm are naturally linear under mass balance.
9.3 Food and nutrition blending
Protein, fat, carbohydrate, and micronutrient constraints are typically linear when expressed per unit mass.
Allergen thresholds and contaminant maximums are also linear constraints when represented as mass fraction limits.
10. Quick reference table for constraint patterns
| Requirement | Typical meaning | Linear constraint form | Common pitfall |
|---|---|---|---|
| Total production | Meet demand Q | Sum x_i = Q | Mixing mass and volume units |
| Minimum quality | At least A_min | Sum(a_i x_i) >= A_min * Q | Using wrong basis for a_i |
| Maximum impurity | No more than A_max | Sum(a_i x_i) <= A_max * Q | ppm conversion mistakes |
| Component availability | Inventory cap | x_i <= U_i | Forgetting safety stock |
| Component share | Recipe limits | min_frac * Q <= x_i <= max_frac * Q | Not updating Q consistently |
| Soft specifications | Allow violation with penalty | Constraint with slack s and penalty | Poor scaling and huge penalties |
FAQ
How do I add a new property constraint in a blending LP.
Define the component property values a_{k,i} on a consistent basis, define the required bound, then add A_k_min * Q <= Sum(a_{k,i} x_i) <= A_k_max * Q as one or two linear constraints.
If the property is not additive, replace it with a validated linear proxy or apply piecewise linearization to preserve linear structure.
How can I model ppm specifications correctly.
Convert ppm to mass fraction by dividing by 1,000,000, and keep all terms on a mass basis.
For a maximum ppm constraint, use Sum(a_i x_i) <= ppm_max_fraction * Q where a_i is also in fraction units.
What should I do if the model becomes infeasible after adding specifications.
Check units and basis first, then relax constraints using slack variables with penalties to identify which specifications drive violations.
Add constraints one at a time starting from mass balance and bounds, and confirm component property data are accurate and within expected ranges.
When is LP not enough for blending optimization.
LP is not enough when essential properties are non-additive, when pricing depends nonlinearly on quality, or when you need logical rules such as mutually exclusive component selection.
In those cases, use piecewise linear approximations and MILP, or use nonlinear programming if the physics and economics require it.
How do I keep the model stable and fast at large scale.
Scale data to comparable magnitudes, avoid mixing percent and fraction units, keep the model linear where possible, and document every property basis and conversion.
Use tight but realistic bounds to reduce the feasible region without cutting off viable solutions.