智能设计
生成式AI算法
GAN、Diffusion Model等生成式模型用于建筑方案自动生成与优化
查看算法细节
from diffusers import StableDiffusionControlNetPipeline
q(x_t|x_{t-1}) = N(x_t; √(1-β_t)x_{t-1}, β_t·I)
p_θ(x_{t-1}|x_t) = N(x_{t-1}; μ_θ(x_t,t), Σ_θ(x_t,t))
pipe = StableDiffusionControlNetPipeline.from_pretrained(
"stabilityai/stable-diffusion-2-1",
controlnet=ControlNetModel.from_pretrained(
"lllyasviel/sd-controlnet-scribble"
)
)
智能设计
强化学习
多目标参数化设计决策，实现空间布局最优解搜索
查看算法细节
from ray.rllib.algorithms import PPO
∇_θ J(π_θ) = E_{s~d_π, a~π_θ}[∇_θ log π_θ(a|s) · A^π(s,a)]
V^π(s) = E_{a~π(s)}[r(s,a) + γ·E_{s'~P(s')}[V^π(s')]]
def reward_function(layout):
return (luminosity * 0.3
+ ventilation * 0.3
+ energy_eff * 0.2
+ cost_efficiency * 0.2)
环境模拟
CFD神经网络代理
计算流体动力学神经网络代理模型，快速预测室内气流分布
查看算法细节
import neuralop.models as nop
∂u/∂t + (u·∇)u = -∇p/ρ + ν∇²u + f
∇·u = 0
model = nop.UNet3d(
input_channels=6,
output_channels=4,
hidden_channels=64
)
model.load_state_dict(torch.load("cfd_proxy_v1.pt"))
环境模拟
物理信息神经网络
PINN技术融合物理规律，实现热舒适度精确预测
查看算法细节
import deepxde as dde
PMV = (0.303·e^(-0.036M) + 0.028)·[(M-W) - 3.05·10⁻³
·(5733 - 6.99(M-W) - Pa) - L·(t_cl - 35.7)]
model = nop.FNO(
n_modes=[16, 16],
hidden_channels=32
)
def pinn_loss(x, y):
return dde.soft_plus(res_thermal(x))
环境模拟
时空序列预测
室内温湿度、空气质量等环境指标的长期预测分析
查看算法细节
from models import Informer
X_enc = PositionalEncoding(Queries, d_model)
X_dec = PositionalEncoding(Keys, d_model)
model = Informer(
enc_in=7,
dec_in=7,
c_out=4,
seq_len=168,
label_len=48,
pred_len=24
)
感知数据
多模态大模型
视觉-语言融合模型，实现设计意图理解与智能问答
查看算法细节
from llava.model import LLaVA
Attention(Q,K,V) = softmax(QK^T/√d_k)·V
Q = X_v · W_q, K = X_l · W_k, V = X_l · W_v
vision_tower = CLIPVisionModel.from_pretrained(
"openai/clip-vit-large-patch14"
)
llm = Vicuna.from_pretrained(
"lmsys/vicuna-7b-v1.5"
)
感知数据
RAG检索增强
规范标准智能问答，政策文件精准解读与检索
查看算法细节
from langchain import RetrievalQA
vectorstore = Milvus.from_documents(
documents=splits,
embedding=OpenAIEmbeddings(
"text-embedding-3-large"
)
)
qa_chain = RetrievalQA.from_chain_type(
llm=ChatGPT("gpt-4-turbo"),
retriever=vectorstore.as_retriever(),
chain_type="stuff"
)
感知数据
点云处理网络
实景三维重建，快速还原真实建筑空间结构
查看算法细节
import pointnet_ops as pnp
model = PointNetSetAbstraction(
npoint=512,
radius=0.2,
nsample=32,
in_channel=6,
mlp=[64,64,128,256]
)
F_θ(r) = Σ_i T_i · (1-exp(-σ_i·δ_i)) · c_i
T_i = exp(-Σ_{j
大模型体系
领域垂类大模型
住房品质评价专用模型微调与优化
查看算法细节
from peft import LoraConfig, get_peft_model
lora_config = LoraConfig(
r=16,
lora_alpha=32,
target_modules=["q_proj", "v_proj"],
lora_dropout=0.05,
bias="none",
task_type=TaskType.CAUSAL_LM
)
base_model = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen2-7B-Instruct",
quantization_config=BitsAndBytesConfig(load_in_4bit=True)
)
大模型体系
Agent智能体框架
多任务协同设计智能体，自动完成复杂设计流程
查看算法细节
from langchain.agents import AgentExecutor
tools = [DesignGenerator(),
CFDSimulator(),
EnergyAnalyzer(),
NormChecker()]
Thought: 思考当前状态
Action: 选择工具
Observation: 观察结果
→ 循环迭代直到完成
agent = create_react_agent(llm, tools)
executor = AgentExecutor(agent=agent, tools=tools)
大模型体系
Benchmark评估体系
多维指标量化评估，领域智能体跑分标准
查看算法细节
import lm_eval
task_dict = {
"housing_safety": {...},
"energy_efficiency": {...},
"design_quality": {...},
"code_compliance": {...}
}
Score = w₁·S_safety + w₂·S_energy + w₃·S_design + w₄·S_comply
w = softmax(importance_scores)
results = lm_eval.evaluate(
model=HousingAgent(),
tasks=task_dict.values()
)
大模型体系
图神经网络
建筑能量流模拟，复杂系统关联分析
查看算法细节
import torch_geometric as pyg
h_v^{(l+1)} = UPDATE(h_v^{(l)}, AGG({h_u^{(l)}, ∀u∈N(v)}))
AGG = Σ_{u∈N(v)} Message(h_u^{(l)}, h_v^{(l)})
model = pyg.GATv2Conv(
in_channels=64,
out_channels=32,
heads=4,
dropout=0.1
)
graph = Data(x=node_features, edge_index=connections)
output = model(graph.x, graph.edge_index)
